General HEVC video steganalysis method based on dual-stream fusion network

Abstract

This invention discloses a general HEVC video steganalysis method based on a dual-stream fusion network, comprising: acquiring compressed video and preprocessing the data to obtain a prediction residual difference map and an intra-frame prediction mode difference map corresponding to each I-frame, and constructing a training set together with frame-level category labels; constructing a general video steganalysis network composed of a dual-stream fusion network, a binary classifier, and a multi-classifier, wherein the dual-stream fusion network takes the two difference maps as inputs to two branches, performs preprocessing, feature extraction, and feature enhancement sequentially, and finally fuses the features to generate a fused feature map, obtains a frame-level feature vector through global average pooling, and the binary classifier outputs the determination result of whether the I-frame is a steganalysis frame; training the network using a domain adversarial training algorithm; and obtaining the video-level determination result through aggregation rules during testing. The advantages are that it can improve the detection accuracy and versatility for multiple steganalysis embedding domains and multiple steganalysis algorithms, especially showing superior performance in steganalysis embedding domain mismatch scenarios.

Description

Technical Field

[0001] This invention belongs to the field of information security technology and relates to a video steganalysis method, particularly a general HEVC (High Efficiency Video Coding) video steganalysis method based on a dual-stream fusion network. Background Technology

[0002] With the widespread adoption of the internet, information security has become increasingly serious. Information hiding techniques, primarily focusing on steganography and steganalysis, are an important research direction in the field of information security. Steganography uses digital multimedia data such as images, audio, and video as carriers to embed secret information in a way that does not cause perceptible changes. Steganalysis, as a countermeasure against steganography, is used to determine whether digital multimedia data contains secret information and further decipher the location and content of that secret information.

[0003] With the widespread adoption of video platforms, video has become a mainstream method of information dissemination in today's society. HEVC has revolutionized various compression technologies, meeting the demands for high definition and fast transmission, while also providing richer and more diverse hiding spaces for video steganography. However, compared to steganalysis in the text and image domains, due to the complexity of H.265 / HEVC encoding, existing dedicated video steganalysis methods all assume that the steganalytic embedding domain is known. But in practical applications, steganalysts often cannot know the specific embedding domain. Therefore, developing video steganalysis methods capable of detecting multiple embedding domains is of great significance. However, multi-domain video steganalysis faces different carrier characteristics and steganalytic distortion effects, and related general video steganalysis work is still in its early stages.

[0004] In recent years, the rapid development of deep learning technology has provided new tools for video steganalysis. Compared to the complexity of manually extracting features, deep learning technology can directly learn the embedding patterns and feature representations of secret information from compressed video data, resulting in higher steganalysis performance. However, most existing research focuses on building dedicated detection models for specific embedding domains (such as motion vectors, predicted partitioning patterns, and transform coefficients). These models typically lack cross-domain generalization ability and are difficult to adapt to multiple steganalysis algorithms simultaneously, thus failing to form a universal video steganalysis method. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a general HEVC video steganalysis method based on a dual-stream fusion network, which can improve the detection accuracy and versatility of various steganalysis embedding domains and various steganalysis algorithms, and has superior performance, especially in steganalysis embedding domain mismatch scenarios.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: a general HEVC video steganalysis method based on a dual-stream fusion network, comprising the following steps: S1. Acquire multiple compressed videos and perform data preprocessing on each compressed video to obtain the prediction residual difference map and intra-frame prediction mode difference map corresponding to each I-frame in each compressed video; combine the prediction residual difference map and intra-frame prediction mode difference map corresponding to all I-frames in the multiple compressed videos with the frame-level category label corresponding to each I-frame to form a training set, wherein the frame-level category label includes a binary classification label indicating whether the I-frame is a steganographic frame, and an embedding domain label indicating the steganographic embedding domain to which the steganographic operation belongs when the I-frame is a steganographic frame. S2. Construct a deep neural network as a general video steganalysis network, which includes a two-stream fusion network, a binary classifier, and a multi-classifier; The dual-stream fusion network is configured as follows: for each I-frame in each compressed video, the prediction residual difference map corresponding to the I-frame is used as the input of the first branch, and the intra-prediction mode difference map corresponding to the I-frame is used as the input of the second branch; the first branch and the second branch respectively perform preprocessing, feature extraction and feature enhancement on their respective inputs to generate a first enhanced feature map and a second enhanced feature map corresponding to the I-frame; the first enhanced feature map and the second enhanced feature map are fused to generate a fused feature map corresponding to the I-frame; The fused feature map is then subjected to global average pooling in the spatial dimension to obtain the frame-level feature vector corresponding to the I-frame for steganalysis. The binary classifier is used to output a determination result of whether the I-frame is a steganographic frame based on the frame-level feature vector. The multi-classifier is used during the training phase to output the determination result of the steganalysis embedding domain to which the steganalysis operation in the steganalysis frame belongs, based on the frame-level feature vector, when the I-frame is a steganalysis frame. S3. Using the training set, train the general video steganalysis network through the domain adversarial training algorithm to obtain the trained general video steganalysis network. S4. For the test compressed video, the same method as in step S1 is used to preprocess the test compressed video to obtain the prediction residual difference map and intra-frame prediction mode difference map corresponding to each I-frame, and input them into the trained general video steganalysis network to obtain the determination result of whether each I-frame is a steganalysis frame; then, based on the determination results of all I-frames, the final determination result of whether the test compressed video is a compressed steganalysis video is obtained through aggregation rules.

[0007] The data preprocessing includes: In the recompression stage, each compressed video is compressed again using the same encoding parameters as its original encoding to obtain a recompressed video; In the data extraction stage, each compressed video and its corresponding recompressed video are decoded to extract all I-frames, and the prediction residual map and intra-frame prediction mode map corresponding to each I-frame are obtained. In the data differencing stage, the prediction residual maps of each compressed video and the corresponding I-frame at the same position in the recompressed video are subtracted and the absolute value is taken to obtain the corresponding prediction residual difference map; the intra-frame prediction mode maps of the I-frame at the same position are subtracted and the absolute value is taken to obtain the corresponding intra-frame prediction mode difference map.

[0008] The dual-stream fusion network adopts a dual-stream convolutional residual network architecture, including the first branch, the second branch, and the dual-stream fusion module; The first branch and the second branch have the same structure, both consisting of a preprocessing module, a feature extraction module, and a feature enhancement module connected in sequence; the preprocessing module, the feature extraction module, and the feature enhancement module all contain convolutional layers, and the feature extraction module and the feature enhancement module all contain residual connections; The first branch inputs the prediction residual difference map, which is processed sequentially by the preprocessing module, the feature extraction module, and the feature enhancement module to output the first enhanced feature map; the second branch inputs the intra-frame prediction mode difference map, which is processed sequentially by the preprocessing module, the feature extraction module, and the feature enhancement module to output the second enhanced feature map; the dual-stream fusion module inputs the first enhanced feature map and the second enhanced feature map, performs feature fusion, and outputs the fused feature map.

[0009] The preprocessing module includes a first convolutional layer, a first batch normalization layer, a first activation function layer, a second convolutional layer, a second batch normalization layer, and a second activation function layer connected in sequence. The first convolutional layer in the first branch of the preprocessing module inputs the prediction residual difference map, the first convolutional layer in the second branch of the preprocessing module inputs the intra-frame prediction mode difference map, and the second activation function layer outputs the corresponding preprocessed feature map.

[0010] The feature extraction module includes a first bottleneck block and a second bottleneck block with the same structure. Both of them include a third batch normalization layer, a third activation function layer, a third convolutional layer, a fourth batch normalization layer, a fourth activation function layer, and a first depthwise separable convolutional layer connected in sequence. The first bottleneck block inputs the preprocessed feature map, and the input of the first bottleneck block is connected with the output residual of the first depth separable convolutional layer in the first bottleneck block. After element-wise addition, a first shallow feature map is generated. The second bottleneck block is input to the first shallow feature map, and the input of the second bottleneck block is connected to the output residual of the first depth separable convolutional layer in the second bottleneck block. After element-wise addition, a second shallow feature map is generated.

[0011] The feature enhancement module includes a first moving inverted bottleneck convolutional block, a second moving inverted bottleneck convolutional block, and a third moving inverted bottleneck convolutional block with the same structure. All three include a fourth convolutional layer, a fifth batch normalization layer, a fifth activation function layer, a second depthwise separable convolutional layer, a sixth batch normalization layer, a sixth activation function layer, channel attention, a first spatial attention, a fifth convolutional layer, and a seventh batch normalization layer. The first moving inverted bottleneck convolutional block is input into the second shallow feature map, and the input of the first moving inverted bottleneck convolutional block is connected to the output residual of the seventh batch normalization layer in the first moving inverted bottleneck convolutional block. After element-wise addition, a first-level enhanced feature map is generated. The second moving inverted bottleneck convolutional block is input into the first-level enhanced feature map, and the input of the second moving inverted bottleneck convolutional block is connected with the output residual of the seventh batch normalization layer in the second moving inverted bottleneck convolutional block. After element-wise addition, a second-level enhanced feature map is generated. The third moving inversion bottleneck convolutional block is input to the second-level enhanced feature map, and the input of the third moving inversion bottleneck convolutional block is connected to the output residual of the seventh batch normalization layer in the third moving inversion bottleneck convolutional block. After element-wise addition, a third-level enhanced feature map is generated. For the feature enhancement module of the first branch, the third-level enhanced feature map is used as the first enhanced feature map, and for the feature enhancement module of the second branch, the third-level enhanced feature map is used as the second enhanced feature map.

[0012] The dual-stream fusion module includes a first convolutional block and a second convolutional block with identical structures, as well as a fusion block. The first convolutional block takes the first enhanced feature map as input and outputs a first convolutional feature map. The second convolutional block takes the second enhanced feature map as input and outputs a second convolutional feature map. The second enhanced feature map is multiplied element-wise with the first convolutional feature map to generate a first weighted feature map. The first enhanced feature map is multiplied element-wise with the second convolutional feature map to generate a second weighted feature map. The first enhanced feature map is added element-wise with the first weighted feature map to generate a first intermediate fusion result. The second enhanced feature map is added element-wise with the second weighted feature map to generate a second intermediate fusion result. The fusion block takes the first intermediate fusion result and the second intermediate fusion result as input and outputs the fused feature map. The first convolutional block and the second convolutional block each include a sixth convolutional layer, a seventh activation function layer, a seventh convolutional layer, and an eighth activation function layer connected in sequence; the fusion block includes an eighth convolutional layer, an eighth batch normalization layer, a ninth activation function layer, a second spatial attention layer, a third depthwise separable convolutional layer, a ninth batch normalization layer, and a tenth activation function layer connected in sequence.

[0013] The aggregation rule is as follows: count the number of I-frames in the test compressed video that are determined to be steganographic frames. If the number exceeds a preset threshold, the test compressed video is determined to be a compressed steganographic video.

[0014] The domain adversarial training algorithm includes: Minimum batches of samples are sampled from the source and target domains, and steganalysis loss and domain discrimination loss are calculated through forward propagation. Each sample includes a prediction residual difference map and an intra-frame prediction mode difference map corresponding to an I-frame. The steganalysis loss is used to measure the difference between the judgment result output by the binary classifier and the corresponding binary classification label, and the domain discrimination loss is used to measure the difference between the judgment result output by the multi-classifier and the corresponding embedded domain label. The network parameters of the binary classifier are updated using the steganalysis classification loss, and the network parameters of the multi-classifier are updated using the domain discrimination loss; the network parameters of the two-stream fusion network are updated using the total loss; wherein, the total loss is the weighted difference between the steganalysis classification loss and the domain discrimination loss, and the gradient of the domain discrimination loss is backpropagated to the two-stream fusion network after being processed by the gradient inversion layer during backpropagation.

[0015] The steganalytic embedding domain determined by the multi-classifier includes either the intra-frame prediction mode domain or the block partitioning structure domain.

[0016] Compared with the prior art, the advantages of the present invention are as follows: By simultaneously processing the prediction residual difference map and the intra-frame prediction mode difference map through a dual-stream fusion network, the general video steganalysis network can learn steganalytic traces from residual information from two different sources, thereby achieving general detection of multiple steganalytic embedding domains and solving the technical problem of existing methods assuming that the steganalytic embedding domains are known.

[0017] By training a general video steganalysis network using a domain adversarial training algorithm, the features learned by the dual-stream fusion network are made invariant across different steganalysis embedding domains. This allows the network to maintain high detection accuracy even in scenarios where steganalysis embedding domains are mismatched, overcoming the shortcomings of existing models that lack cross-domain generalization ability.

[0018] By designing frame-level feature vectors and aggregation rules, the judgment results of all I-frames in the video can be comprehensively compressed to make a final video-level decision, thereby avoiding the impact of single-frame misjudgment on the overall result and improving the reliability of detection.

[0019] During data preprocessing, the compressed video is recompressed and then differentially compared with the original compressed video. This effectively amplifies the subtle modifications caused by steganography, making it easier for the subsequent dual-stream fusion network to capture the steganography traces and improving the detection sensitivity.

[0020] The dual-stream fusion network sets up two independent branches to process the prediction residual difference map and the intra-frame prediction mode difference map respectively. This enables the dual-stream fusion network to perform parallel feature extraction for two different types of steganalysis, thereby avoiding mutual interference between features and improving the targeting and effectiveness of feature extraction.

[0021] The dual-stream fusion network consists of four stages: preprocessing, feature extraction, feature enhancement, and feature fusion. Residual connections are incorporated into the feature extraction and enhancement modules to allow the network to be trained deeper without gradient vanishing, thus enabling it to learn more subtle and abstract steganalytic features. The preprocessing module uses two consecutive convolutional layers with batch normalization and ReLU activation to effectively extract preliminary features from the input difference map, providing high-quality initial feature representations for subsequent feature extraction and enhancement stages. The bottleneck block in the feature extraction module combines convolutional layers and depthwise separable convolutional layers, significantly reducing the number of parameters and computational cost while maintaining feature extraction capabilities, thereby improving the processing efficiency of the dual-stream fusion network and reducing computational resource requirements during deployment. In the feature extraction stage, the concatenation of the two bottleneck blocks and residual connections enable the dual-stream fusion network to extract deeper feature representations, thereby improving its ability to model complex steganalytic artifacts. The proposed moving-inverted bottleneck convolutional block in the feature enhancement module utilizes a dual mechanism of channel attention and spatial attention, enabling the two-stream fusion network to automatically enhance steganalytic-related feature responses and suppress irrelevant noise, thereby improving the signal-to-noise ratio and detection accuracy. In the feature enhancement stage, the cascaded structure of three moving-inverted bottleneck convolutional blocks progressively enhances features, achieving multi-level attention feature extraction from shallow to deep layers and enhancing the richness of feature representations. The proposed two-stream fusion module in the feature fusion stage achieves complementary fusion of information from the two branches, resulting in a fused feature map containing more comprehensive and richer steganalytic information while avoiding the loss of important information during the fusion process.

[0022] By comparing the number of steg frames in all I-frames with a preset threshold, the video-level judgment result integrates information from all I-frames in the video, thereby avoiding the high false detection rate caused by judgment based on a single frame and improving the reliability of the detection result.

[0023] By inverting the gradient of the domain discrimination loss during backpropagation through the gradient inversion layer (GRL), the dual-stream fusion network is forced to extract general steganalytic trace features that are independent of the steganalytic embedding domain, which significantly improves the model's generalization ability in steganalytic embedding domain mismatch scenarios.

[0024] By designing the total loss function as a weighted difference between the steganalysis classification loss and the domain discrimination loss, the network can simultaneously achieve two objectives during the optimization process: minimizing the steganalysis classification error and maximizing the domain discrimination error. This enables end-to-end domain adversarial training without the need for complex multi-stage training procedures. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the overall implementation of the training phase of the method of the present invention. Figure 2 A schematic diagram of the composition structure of the two-stream fusion network in the general video steganalysis network constructed by the method of the present invention; Figure 3 This is a schematic diagram of the composition of the preprocessing module in a dual-stream fusion network; Figure 4 This is a schematic diagram of the composition structure of the bottleneck block in the feature extraction module of a two-stream fusion network. Figure 5 This is a schematic diagram of the composition structure of the Moving Inverse Bottleneck Convolutional Block (MBConv block) in the feature enhancement module of a two-stream fusion network. Figure 6 This is a schematic diagram of the composition structure of the dual-stream fusion module in a dual-stream fusion network. Detailed Implementation

[0026] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0027] This embodiment provides a general HEVC video steganalysis method based on a two-stream fusion network, which includes a training phase and an inference phase, such as... Figure 1 As shown, the training phase includes the following steps: S1. Data Preprocessing and Training Set Construction: Acquire multiple compressed videos and perform data preprocessing on each compressed video to obtain the prediction residual difference map and intra-frame prediction mode difference map corresponding to each I-frame in each compressed video. Combine the prediction residual difference map and intra-frame prediction mode difference map corresponding to all I-frames in the multiple compressed videos with the frame-level category label corresponding to each I-frame to form a training set. The frame-level category label includes a binary classification label indicating whether the I-frame is a steganographic frame (e.g., 0 represents the original frame and 1 represents the steganographic frame), and an embedding domain label indicating the steganographic embedding domain to which the steganographic operation belongs when the I-frame is a steganographic frame (e.g., 0 represents the intra-frame prediction mode domain and 1 represents the block partitioning structure domain).

[0028] In step S1, the obtained compressed video may be the original compressed video obtained by directly encoding with the H.265 / HEVC standard encoder, or it may be the compressed stegated video obtained by encoding with the H.265 / HEVC standard encoder and then performing stegating processing in a certain stegation embedding domain. The stegating processing can use a variety of stegating algorithms, such as stegating methods for the intra-predictive mode domain or stegating methods for the block partitioning structure domain.

[0029] In step S1, data preprocessing includes a recompression stage, a data extraction stage, and a data differentiation stage.

[0030] In the recompression stage, each compressed video is compressed again using the same encoding parameters as its original encoding (including quantization parameters QP, group of pictures (GOP) structure, number of reference frames, encoding profile, etc.) to obtain a recompressed video. The purpose of this step is to calibrate the statistical characteristics differences caused by different video content or encoding history, thereby amplifying the modification traces caused by steganography.

[0031] In the data extraction stage, each compressed video and its corresponding recompressed video are decoded. During the decoding process, all I-frames (intra-coded frames) are extracted, and the prediction residual map and intra-prediction mode map corresponding to each I-frame are obtained. The prediction residual map records the difference between the current frame pixel value and the predicted value, and the intra-prediction mode map records the intra-prediction mode number used by each coding block.

[0032] In the data differencing stage, the prediction residual maps of each compressed video and its corresponding recompressed video at the same position (i.e., the same frame number) are subtracted (i.e., pixel-by-pixel) and the absolute value is taken to obtain the corresponding prediction residual difference map. Similarly, the intra-frame prediction mode maps of the I-frames at the same position are subtracted and the absolute value is taken to obtain the corresponding intra-frame prediction mode difference map. These two difference maps can effectively amplify the faint traces left by steganography operations in the steganography embedding domain, which is beneficial for subsequent detection by deep neural networks.

[0033] After the above data preprocessing, the resulting prediction residual difference map and intra-frame prediction mode difference map are both H×W×1 in size, where H is the height and W is the width.

[0034] S2. Construct a deep neural network as a general video steganalysis network: such as Figure 1 As shown, a general video steganalysis network includes a two-stream fusion network, a binary classifier, and a multi-classifier. Figure 2 As shown, the dual-stream fusion network is configured as follows: for each I-frame in each compressed video, the prediction residual difference map corresponding to the I-frame is used as the input of the first branch, and the intra-prediction mode difference map corresponding to the I-frame is used as the input of the second branch; the first and second branches respectively preprocess, extract features, and enhance features from their respective inputs, generating a first enhanced feature map and a second enhanced feature map corresponding to the I-frame; the first and second enhanced feature maps are fused to generate a fused feature map corresponding to the I-frame. The fused feature map is then subjected to global average pooling in the spatial dimension. Global average pooling compresses the feature map of each channel into a single value, thereby aggregating spatial information into statistical information, resulting in a frame-level feature vector for steganalysis corresponding to the I-frame. A binary classifier is used to output a determination result of whether the I-frame is a steganalysis frame based on the frame-level feature vector. A multi-classifier is used during the training phase to output a determination result of the steganalysis embedding domain to which the steganalysis operation in the steganalysis frame belongs, based on the frame-level feature vector when the I-frame is a steganalysis frame. The steganalysis embedding domain includes an intra-prediction mode domain or a block partitioning structure domain.

[0035] Specifically, such as Figures 2 to 5 As shown, the dual-stream fusion network adopts a dual-stream convolutional residual network architecture, including a first branch, a second branch, and a dual-stream fusion module. The first and second branches have identical structures, both consisting of a preprocessing module, a feature extraction module, and a feature enhancement module connected sequentially. Each of these modules contains convolutional layers, and both the feature extraction and feature enhancement modules include residual connections. The first branch takes the prediction residual difference map as input, processes it sequentially through the preprocessing, feature extraction, and feature enhancement modules, and outputs a first enhanced feature map. The second branch takes the intra-frame prediction mode difference map as input, processes it sequentially through the preprocessing, feature extraction, and feature enhancement modules, and outputs a second enhanced feature map. The dual-stream fusion module takes the first and second enhanced feature maps as input, performs feature fusion, and outputs a fused feature map.

[0036] In this embodiment, the preprocessing module is used to perform preliminary feature extraction on the input difference map, specifically, as follows: Figure 3As shown, the preprocessing module includes a first convolutional layer, a first batch normalization layer, a first activation function layer, a second convolutional layer, a second batch normalization layer, and a second activation function layer connected in sequence. In the first branch of the preprocessing module, the first convolutional layer inputs the prediction residual difference map, and in the second branch, the first convolutional layer inputs the intra-frame prediction mode difference map. The second activation function layer outputs the corresponding preprocessed feature map, which has a size of H×W×16. The kernel size of both the first and second convolutional layers is 3×3, the stride is 1, and the padding method is SAME. The first convolutional layer has 1 input channel and 16 output channels, and the second convolutional layer has 16 input channels and 16 output channels. Both the first and second activation function layers use the ReLU activation function.

[0037] In this embodiment, as Figure 2 As shown, the feature extraction module includes a first bottleneck block (Bottleneck block) and a second bottleneck block with identical structures, as follows: Figure 4 As shown, both layers consist of a third batch normalization layer, a third activation function layer, a third convolutional layer, a fourth batch normalization layer, a fourth activation function layer, and a first depthwise separable convolutional layer, connected sequentially. The first bottleneck block receives the preprocessed feature map as input, and the input of the first bottleneck block is concatenated with the output residual of the first depthwise separable convolutional layer within it. After element-wise addition, a first shallow feature map is generated, with a size of H×W×16. The second bottleneck block receives the first shallow feature map as input, and the input of the second bottleneck block is concatenated with the output residual of the first depthwise separable convolutional layer within it. After element-wise addition, a second shallow feature map is generated, with a size of H×W×16. The third convolutional layer has a 1×1 kernel size for inter-channel information exchange; the first depthwise separable convolutional layer has a 3×3 kernel size, performing convolution operations channel-by-channel to reduce computational cost; both the third and fourth activation function layers use ReLU activation functions.

[0038] In this embodiment, the feature enhancement module is used to enhance important features through an attention mechanism, such as... Figure 2 As shown, the feature enhancement module includes a first moving-inverted bottleneck convolutional block (MBConv block), a second moving-inverted bottleneck convolutional block, and a third moving-inverted bottleneck convolutional block with identical structures, as follows: Figure 5As shown, all three layers include a fourth convolutional layer, a fifth batch normalization layer, a fifth activation function layer, a second depthwise separable convolutional layer, a sixth batch normalization layer, a sixth activation function layer, channel attention, a first spatial attention layer, a fifth convolutional layer, and a seventh batch normalization layer. The first moving-inverted bottleneck convolutional block takes the second shallow feature map as input, and the input of the first moving-inverted bottleneck convolutional block is concatenated with the output residual of the seventh batch normalization layer within the first moving-inverted bottleneck convolutional block. After element-wise addition, a first-level enhanced feature map is generated, with a size of H×W×32. The second moving-inverted bottleneck convolutional block takes the first-level enhanced feature map as input, and the input of the second moving-inverted bottleneck convolutional block is concatenated with the output residual of the seventh batch normalization layer within the second moving-inverted bottleneck convolutional block. After element-wise addition, a second-level enhanced feature map is generated, with a size of H×W×64. The input of the third moving-inverted bottleneck convolutional block to the second-level enhanced feature map is concatenated with the output residual of the seventh batch normalization layer within the third moving-inverted bottleneck convolutional block. After element-wise addition, a third-level enhanced feature map is generated, with a size of H×W×64. For the feature enhancement module of the first branch, the third-level enhanced feature map serves as the first enhanced feature map; for the feature enhancement module of the second branch, it serves as the second enhanced feature map. The kernel size of the fourth and fifth convolutional layers is 1×1, used for dimensionality upscaling and downscaling, respectively; the kernel size of the second depthwise separable convolutional layer is 3×3; and the fifth and sixth activation function layers both use the Swish activation function.

[0039] In this embodiment, channel attention is configured as follows: the input feature map is sequentially subjected to global average pooling, multilayer perceptron mapping, and Softmax function normalization to generate channel weights. The channel weights are then multiplied channel-by-channel with the input feature map to obtain a channel attention feature map. The sizes of the channel attention feature maps output by the channel attention in the first, second, and third moving-inverted bottleneck convolutional blocks are H×W×32, H×W×64, and H×W×128, respectively. First spatial attention is configured as follows: the input feature map is subjected to global average pooling in both the horizontal and vertical directions. The pooling results in both directions are concatenated and then sequentially subjected to convolution and nonlinear transformation. After separation, the results are processed by the Sigmoid function to generate two directional weights. Finally, the two directional weights are multiplied by the input feature map, and the two multiplication results are added to obtain a spatial attention feature map. The sizes of the spatial attention feature maps output by the first spatial attention in the first, second, and third moving-inverted bottleneck convolutional blocks are H×W×32, H×W×64, and H×W×128, respectively.

[0040] In this embodiment, as Figure 6As shown, the dual-stream fusion module includes a first convolutional block and a second convolutional block with identical structures, as well as a fusion block. The first convolutional block takes a first enhanced feature map as input and outputs a first convolutional feature map with a size of H×W×64. The second convolutional block takes a second enhanced feature map as input and outputs a second convolutional feature map with a size of H×W×64. The second enhanced feature map is multiplied element-wise with the first convolutional feature map to generate a first weighted feature map. The first enhanced feature map is multiplied element-wise with the second convolutional feature map to generate a second weighted feature map. The first enhanced feature map is added element-wise with the first weighted feature map to generate a first intermediate fusion result. The second enhanced feature map is added element-wise with the second weighted feature map to generate a second intermediate fusion result. The fusion block takes the first intermediate fusion result and the second intermediate fusion result as input and outputs a fused feature map with a size of H×W×128.

[0041] Among them, such as Figure 6 As shown, both the first and second convolutional blocks include a sixth convolutional layer, a seventh activation function layer, a seventh convolutional layer, and an eighth activation function layer connected in sequence. The fusion block includes an eighth convolutional layer, an eighth batch normalization layer, a ninth activation function layer, a second spatial attention layer, a third depthwise separable convolutional layer, a ninth batch normalization layer, and a tenth activation function layer connected in sequence. The fusion block first concatenates the first intermediate fusion result and the second intermediate fusion result in the channel dimension, and the concatenated feature map is input to the eighth convolutional layer. The kernel size of the sixth, seventh, and eighth convolutional layers is 1×1, and the kernel size of the third depthwise separable convolutional layer is 3×3. The seventh, ninth, and tenth activation function layers are all ReLU activation functions, and the eighth activation function layer is a Sigmoid activation function. The first and second convolutional feature maps are both weight maps output by the Sigmoid activation function, with values ​​in the range of (0,1). The configuration of the second spatial attention is the same as that of the first spatial attention.

[0042] In this embodiment, the size of the frame-level feature vector is 128×1.

[0043] In this embodiment, the binary classifier is implemented by a fully connected layer followed by a Softmax activation function, and outputs a two-dimensional probability vector, which represents the probability that the I-frame is an original frame or a steganalytic frame, and the larger probability is taken as the judgment result.

[0044] In this embodiment, the multi-classifier is implemented by a fully connected layer followed by a Softmax activation function, and outputs a two-dimensional probability vector, which represents the probability that the steganography operation of the steganography frame belongs to the intra-predictive mode domain or the block partitioning structure domain, respectively, and the larger probability is taken as the judgment result.

[0045] S3. Training the network: Using the training set, the general video steganalysis network is trained through the domain adversarial training algorithm to obtain the trained general video steganalysis network.

[0046] In this embodiment, the domain adversarial training algorithm adopts the domain adversarial neural network (DANN) framework, specifically including: 1) Sample the minimum batch of samples from the source and target domains, and calculate the steganalysis loss and domain discrimination loss through forward propagation. Each sample includes a prediction residual difference map and an intra-frame prediction mode difference map corresponding to an I-frame. The steganalysis loss is used to measure the difference between the judgment result output by the binary classifier and the corresponding binary classification label. The domain discrimination loss is used to measure the difference between the judgment result output by the multi-classifier and the corresponding embedded domain label. Both steganalysis loss and domain discrimination loss adopt cross-entropy loss.

[0047] In this embodiment, the source domain consists of samples with frame-level category labels under one steganalysis embedding domain (e.g., intra-prediction mode domain) in the training set, including original samples (i.e., samples composed of the prediction residual difference map and the intra-prediction mode difference map corresponding to the I-frame in the compressed original video) and steganalysis samples (i.e., samples composed of the prediction residual difference map and the intra-prediction mode difference map corresponding to the I-frame in the compressed steganalysis video); the target domain consists of steganalysis samples with frame-level category labels under another steganalysis embedding domain (e.g., block partitioning structure domain). The goal of domain adversarial training is to make the features extracted by the two-stream fusion network invariant to different steganalysis embedding domains, so that the classifier trained on the source domain can be directly applied to the target domain.

[0048] 2) Update the network parameters of the binary classifier using steganalysis classification loss through backpropagation, and update the network parameters of the multi-classifier using domain discrimination loss through backpropagation; update the network parameters of the two-stream fusion network using total loss through backpropagation; where total loss is the weighted difference between steganalysis classification loss and domain discrimination loss, i.e., total loss = steganalysis classification loss - weight × domain discrimination loss, where weight is a balancing hyperparameter, such as 0.1. During backpropagation, the gradient of domain discrimination loss is processed by gradient inversion layer (GRL) and then backpropagated to the two-stream fusion network.

[0049] In this embodiment, the training process continues for multiple iterations until the total loss converges.

[0050] Inference phase: For the test compressed video, the same method as in step S1 is used to preprocess the test compressed video to obtain the prediction residual difference map and intra-frame prediction mode difference map corresponding to each I-frame, and input them into the trained general video steganalysis network to obtain the determination result of whether each I-frame is a steganalysis frame; then, based on the determination results of all I-frames, the final determination result of whether the test compressed video is a compressed steganalysis video is obtained through aggregation rules.

[0051] In step S4, the test compressed video may be the original compressed video obtained by directly encoding with an H.265 / HEVC standard encoder, or it may be a compressed steganographic video obtained by encoding with an H.265 / HEVC standard encoder and then performing steganographic embedding domain steganography.

[0052] In this embodiment, the aggregation rule is as follows: count the number of I-frames in the test compressed video that are determined to be steganographic frames. If this number exceeds a preset threshold, the test compressed video is determined to be a compressed steganographic video. As a preferred example, the preset threshold can be set to half of the total number of I-frames in the test compressed video, that is, a majority voting mechanism is adopted: if more than half of the I-frames are determined to be steganographic frames, the entire compressed video is determined to be a compressed steganographic video; otherwise, it is determined to be a compressed original video.

[0053] To further illustrate the feasibility and effectiveness of the method of the present invention, simulation experiments were conducted on the method of the present invention.

[0054] In the simulation experiments, the video sequences used were all from the Internet (https: / / media.xiph.org / video / derf / ), containing 116 4:2:0 YUV video sequences with a resolution of 288×352, and an average of 261 frames per YUV video sequence. The HEVC reference software HM 16.15 was used to encode the YUV video sequences, where the GOP was an IPPP structure, and the encoding configuration file was encoder_lowdelay_P_main.cfg.

[0055] In the simulation experiment, three different steganography algorithms in the intra-prediction mode domain and three different steganography algorithms in the block partitioning structure domain were used to perform steganography processing on the encoded video sequence to obtain compressed steganography video. Then, four different steganography analysis methods were used to detect the compressed steganography video.

[0056] Three different steganography algorithms in the intra-prediction mode domain are: Tar1 (from reference 1, Y. Dong, T. Sun and X. Jiang, “A High Capacity HEVC Steganographic Algorithm Using IntraPrediction Modes in Multi-sized Prediction Blocks,” in 17th International Workshop on Digital Watermarking (IWDW), pp. 233–247, Oct. 2018, Jeju Island, Korea), and the steganography method (referred to as Tar1) from reference 2 (Y. Dong, X. Jiang, Z. Li, et al., “Multi-channel HEVC steganography by minimizing IPM steganographic distortions,” IEEE Transactions on Multimedia). The steganography method (referred to as Tar2) from reference 3 (L. Yang, D. Xu, J. Qian, et al. “IPM Priority-Preserving Adaptive Steganography for HEVC,” IEEE Transactions on Circuits and Systems for Video Technology, 2026, 36(1): 537-550.) is a steganography method (referred to as Tar3).

[0057] The three different steganography algorithms within the block partitioning structure domain are derived from the steganography method in reference 4 (Y. Tew and K. Wong, “Information hiding in HEVC standard using adaptive coding block size decision,” IEEE International Conference on Image Processing (ICIP). 2014, pp. 5502–5506. ) (referred to as Tar4), the steganography method in reference 5 (Y. Dong, X. Jiang, Z. Li, “Adaptive HEVC Steganography Based on Steganographic Compression Efficiency Degradation Model,” IEEE Transactions on Dependable and Secure Computing, vol. 20, no. 1, pp. 769–783, Jan. 2023. ) (referred to as Tar5), and the steganography method in reference 6 (L. Yang, D. Xu, J. Qian, et al. ) (referred to as Tar5). al. "Quad-tree structure-preserving adaptive steganography for HEVC." IEEE Transactions on Multimedia, vol.26, pp. 8625-8638, 2024. This refers to a steganography method (abbreviated as Tar6) based on quad-tree structure-preserving adaptive HEVC.

[0058] The four different steganalysis methods are as follows: Zhao (Y. Zhao, H. Zhang, Y. Cao, et al. “Video Steganalysis Based on Intra Prediction Mode Calibration.” International Workshop on Digital Watermarking, Tokyo, Japan, 2015: 119-113.), Cao (M. Cao, L. Tian, ​​C. Li. “A Steganalytic Approach to Detect Intra Prediction Mode Modification Using Difference of Partitioning Structure for HEVC.” IEEE Transactions on Consumer Electronics, 2025, 71(2): 2702-2713.), Cao (L. Zhai, L. Wang, Y. Ren. “Universal Detection of Video Steganography in Multiple Domains Based on…” ...L. Zhai, L. Wang, Y. Ren. “Universal Detection of Video Steganography in Multiple Domains Based on…”), Cao (L. Zhai, L. Wang, Y. Ren. “Universal Detection of Video Steganography in Multiple Domains Based on…”), Cao (L. Zhai, L. Wang, Y. Ren. “Universal Detection of Video Steganography in Multiple Domains Based on…”), Cao (L. Zhai, L. Wang, Y. Ren. “Universal Detection of Video The Consistency of Motion Vectors.” IEEE Transactions on Information Forensics and Security, 2020, 15: 1762-1777. (General detection of multi-domain video steganography based on motion vector consistency, IEEE Transactions on Information Forensics and Security) (abbreviated as Zhai), and reference 10 (H. Dai, D. Xu, L. Yang, et al. “HEVC Video Steganalysis Based on Centralized Error and Attention Mechanism. IEEE Transactions on Multimedia, 2025, 27: 8914-8925. (HEVC video steganography analysis based on centralized error and attention mechanism, IEEE Multimedia Journal)) (abbreviated as Dai).

[0059] In the simulation experiments, accuracy (Acc) was used as the evaluation index. The higher the value of Acc, the stronger the detection performance. Table 1 shows a comparison of the accuracy (Acc) of training on Tar1, Tar2, and Tar3 and testing on Tar4, Tar5, and Tar6 using Zhao, Cao, and the method of this invention, respectively. Table 2 shows a comparison of the accuracy (Acc) of training on Tar4, Tar5, and Tar6 and testing on Tar1, Tar2, and Tar3 using Zhai, Dai, and the method of this invention, respectively.

[0060] Table 1 shows the accuracy (Acc) achieved by Zhao, Cao, and the method of this invention when training on Tar1, Tar2, and Tar3 and testing on Tar4, Tar5, and Tar6, respectively.

[0061] Table 2 shows the accuracy (Acc) of training on Tar4, Tar5, and Tar6 and testing on Tar1, Tar2, and Tar3 using Zhai, Dai, and the method of this invention, respectively.

[0062] As shown in Tables 1 and 2, the method of this invention significantly outperforms the methods of Zhao and Cao in mismatch scenarios for intra-prediction mode domain and block partitioning structure domain detection. For general detection scenarios involving both intra-prediction mode domain and block partitioning structure domain, the method of this invention also exhibits significant advantages over the methods of Zhao and Dai. Therefore, the method of this invention demonstrates higher detection performance for compressed stegographic videos with various steganalysis embedding domains and possesses more general applicability in steganalysis embedding domain mismatch scenarios.

Claims

1. A general HEVC video steganalysis method based on a dual-stream fusion network, characterized in that, Includes the following steps: S1. Acquire multiple compressed videos and perform data preprocessing on each compressed video to obtain the prediction residual difference map and intra-frame prediction mode difference map corresponding to each I-frame in each compressed video. The prediction residual difference map and intra-prediction mode difference map corresponding to all I-frames in multiple compressed videos, together with the frame-level category label corresponding to each I-frame, constitute a training set. The frame-level category label includes a binary classification label indicating whether the I-frame is a steganographic frame, and an embedding domain label indicating the steganographic embedding domain to which the steganographic operation belongs when the I-frame is a steganographic frame. S2. Construct a deep neural network as a general video steganalysis network, which includes a two-stream fusion network, a binary classifier, and a multi-classifier; The dual-stream fusion network is configured as follows: for each I-frame in each compressed video, the prediction residual difference map corresponding to the I-frame is used as the input of the first branch, and the intra-prediction mode difference map corresponding to the I-frame is used as the input of the second branch; the first branch and the second branch respectively perform preprocessing, feature extraction and feature enhancement on their respective inputs to generate a first enhanced feature map and a second enhanced feature map corresponding to the I-frame; the first enhanced feature map and the second enhanced feature map are fused to generate a fused feature map corresponding to the I-frame; The fused feature map is then subjected to global average pooling in the spatial dimension to obtain the frame-level feature vector corresponding to the I-frame for steganalysis. The binary classifier is used to output a determination result of whether the I-frame is a steganographic frame based on the frame-level feature vector. The multi-classifier is used during the training phase to output the determination result of the steganalysis embedding domain to which the steganalysis operation in the steganalysis frame belongs, based on the frame-level feature vector, when the I-frame is a steganalysis frame. S3. Using the training set, train the general video steganalysis network through the domain adversarial training algorithm to obtain the trained general video steganalysis network. S4. For the test compressed video, the same method as in step S1 is used to preprocess the test compressed video to obtain the prediction residual difference map and intra-frame prediction mode difference map corresponding to each I-frame, and input them into the trained general video steganalysis network to obtain the determination result of whether each I-frame is a steganalysis frame; then, based on the determination results of all I-frames, the final determination result of whether the test compressed video is a compressed steganalysis video is obtained through aggregation rules.

2. The general HEVC video steganalysis method based on a dual-stream fusion network according to claim 1, characterized in that, The data preprocessing includes: In the recompression stage, each compressed video is compressed again using the same encoding parameters as its original encoding to obtain a recompressed video; In the data extraction stage, each compressed video and its corresponding recompressed video are decoded to extract all I-frames, and the prediction residual map and intra-frame prediction mode map corresponding to each I-frame are obtained. In the data differencing stage, the prediction residual maps of each compressed video and the corresponding I-frame at the same position in the recompressed video are subtracted and the absolute value is taken to obtain the corresponding prediction residual difference map; the intra-frame prediction mode maps of the I-frame at the same position are subtracted and the absolute value is taken to obtain the corresponding intra-frame prediction mode difference map.

3. The general HEVC video steganalysis method based on a dual-stream fusion network according to claim 1 or 2, characterized in that, The dual-stream fusion network adopts a dual-stream convolutional residual network architecture, including the first branch, the second branch, and the dual-stream fusion module; The first branch and the second branch have the same structure, both consisting of a preprocessing module, a feature extraction module, and a feature enhancement module connected in sequence; the preprocessing module, the feature extraction module, and the feature enhancement module all contain convolutional layers, and the feature extraction module and the feature enhancement module all contain residual connections; The first branch inputs the prediction residual difference map, which is processed sequentially by the preprocessing module, the feature extraction module, and the feature enhancement module to output the first enhanced feature map; the second branch inputs the intra-frame prediction mode difference map, which is processed sequentially by the preprocessing module, the feature extraction module, and the feature enhancement module to output the second enhanced feature map; the dual-stream fusion module inputs the first enhanced feature map and the second enhanced feature map, performs feature fusion, and outputs the fused feature map.

4. The general HEVC video steganalysis method based on a dual-stream fusion network according to claim 3, characterized in that, The preprocessing module includes a first convolutional layer, a first batch normalization layer, a first activation function layer, a second convolutional layer, a second batch normalization layer, and a second activation function layer connected in sequence. The first convolutional layer in the first branch of the preprocessing module inputs the prediction residual difference map, the first convolutional layer in the second branch of the preprocessing module inputs the intra-frame prediction mode difference map, and the second activation function layer outputs the corresponding preprocessed feature map.

5. The general HEVC video steganalysis method based on a dual-stream fusion network according to claim 4, characterized in that, The feature extraction module includes a first bottleneck block and a second bottleneck block with the same structure. Both of them include a third batch normalization layer, a third activation function layer, a third convolutional layer, a fourth batch normalization layer, a fourth activation function layer, and a first depthwise separable convolutional layer connected in sequence. The first bottleneck block inputs the preprocessed feature map, and the input of the first bottleneck block is connected with the output residual of the first depth separable convolutional layer in the first bottleneck block. After element-wise addition, a first shallow feature map is generated. The second bottleneck block is input to the first shallow feature map, and the input of the second bottleneck block is connected to the output residual of the first depth separable convolutional layer in the second bottleneck block. After element-wise addition, a second shallow feature map is generated.

6. The general HEVC video steganalysis method based on a dual-stream fusion network according to claim 5, characterized in that, The feature enhancement module includes a first moving inverted bottleneck convolutional block, a second moving inverted bottleneck convolutional block, and a third moving inverted bottleneck convolutional block with the same structure. All three include a fourth convolutional layer, a fifth batch normalization layer, a fifth activation function layer, a second depthwise separable convolutional layer, a sixth batch normalization layer, a sixth activation function layer, channel attention, a first spatial attention, a fifth convolutional layer, and a seventh batch normalization layer. The first moving inverted bottleneck convolutional block is input into the second shallow feature map, and the input of the first moving inverted bottleneck convolutional block is connected to the output residual of the seventh batch normalization layer in the first moving inverted bottleneck convolutional block. After element-wise addition, a first-level enhanced feature map is generated. The second moving inverted bottleneck convolutional block is input into the first-level enhanced feature map, and the input of the second moving inverted bottleneck convolutional block is connected with the output residual of the seventh batch normalization layer in the second moving inverted bottleneck convolutional block. After element-wise addition, a second-level enhanced feature map is generated. The third moving inversion bottleneck convolutional block is input to the second-level enhanced feature map, and the input of the third moving inversion bottleneck convolutional block is connected to the output residual of the seventh batch normalization layer in the third moving inversion bottleneck convolutional block. After element-wise addition, a third-level enhanced feature map is generated. For the feature enhancement module of the first branch, the third-level enhanced feature map is used as the first enhanced feature map, and for the feature enhancement module of the second branch, the third-level enhanced feature map is used as the second enhanced feature map.

7. The general HEVC video steganalysis method based on a dual-stream fusion network according to claim 6, characterized in that, The dual-stream fusion module includes a first convolutional block and a second convolutional block with identical structures, as well as a fusion block. The first convolutional block takes the first enhanced feature map as input and outputs a first convolutional feature map. The second convolutional block takes the second enhanced feature map as input and outputs a second convolutional feature map. The second enhanced feature map is multiplied element-wise with the first convolutional feature map to generate a first weighted feature map. The first enhanced feature map is multiplied element-wise with the second convolutional feature map to generate a second weighted feature map. The first enhanced feature map is added element-wise with the first weighted feature map to generate a first intermediate fusion result. The second enhanced feature map and the second weighted feature map are added element by element to generate the second intermediate fusion result; The fusion block takes the first intermediate fusion result and the second intermediate fusion result as inputs and outputs the fusion feature map. The first convolutional block and the second convolutional block each include a sixth convolutional layer, a seventh activation function layer, a seventh convolutional layer, and an eighth activation function layer connected in sequence; the fusion block includes an eighth convolutional layer, an eighth batch normalization layer, a ninth activation function layer, a second spatial attention layer, a third depthwise separable convolutional layer, a ninth batch normalization layer, and a tenth activation function layer connected in sequence.

8. The general HEVC video steganalysis method based on a dual-stream fusion network according to claim 1, characterized in that, The aggregation rule is as follows: count the number of I-frames in the test compressed video that are determined to be steganographic frames. If the number exceeds a preset threshold, the test compressed video is determined to be a compressed steganographic video.

9. The general HEVC video steganalysis method based on a dual-stream fusion network according to claim 1, characterized in that, The domain adversarial training algorithm includes: Minimum batches of samples are sampled from the source and target domains, and steganalysis loss and domain discrimination loss are calculated through forward propagation. Each sample includes a prediction residual difference map and an intra-frame prediction mode difference map corresponding to an I-frame. The steganalysis loss is used to measure the difference between the judgment result output by the binary classifier and the corresponding binary classification label, and the domain discrimination loss is used to measure the difference between the judgment result output by the multi-classifier and the corresponding embedded domain label. The network parameters of the binary classifier are updated using the steganalysis classification loss, and the network parameters of the multi-classifier are updated using the domain discrimination loss; the network parameters of the two-stream fusion network are updated using the total loss; wherein, the total loss is the weighted difference between the steganalysis classification loss and the domain discrimination loss, and the gradient of the domain discrimination loss is backpropagated to the two-stream fusion network after being processed by the gradient inversion layer during backpropagation.

10. The general HEVC video steganalysis method based on a dual-stream fusion network according to claim 1, characterized in that, The steganalytic embedding domain determined by the multi-classifier includes either the intra-frame prediction mode domain or the block partitioning structure domain.

Images