HEVC transformation coefficient video steganalysis method based on attention mechanism

Abstract

The invention discloses an attention mechanism-based HEVC (High Efficiency Video Coding) transformation coefficient video steganalysis method, which comprises the following steps of: performing data preprocessing on a compressed video, and extracting a prediction residual image corresponding to an I frame to form a training set; then, constructing a deep neural network comprising a convolution sub-network, a self-attention sub-network and a binary classifier, wherein the deep neural network is used for extracting steganography traces and classifying videos; through training and testing, the network can accurately judge whether the video is subjected to steganography processing or not; the steganalysis method has the advantages that steganalysis is carried out in combination with the deep learning technology, the accuracy and robustness of steganalysis can be improved, the steganalysis method can adapt to diversified steganalysis methods, and high detection performance is achieved in a low-embedding-rate scene.

Description

Technical Field

[0001] The present invention relates to a video steganalysis method, in particular to a HEVC transform coefficient video steganalysis method based on an attention mechanism. Background Art

[0002] With the popularization of the Internet and the widespread application of digital media, the way and efficiency of information transmission have undergone qualitative changes, and the situation of information security has become increasingly severe. Information hiding technology, with steganalysis as the main research content, is an important research direction in the field of information security, and has broad application prospects in the military, commercial and industrial fields. Steganography is an information hiding technology that realizes masked communication, embedding secret information in multimedia data in a way that does not cause perceptible changes. Although steganography can improve the security of secret communication, it is also very easy to be used by criminals or hostile forces, causing adverse security impacts. In order to curb the harm caused by the illegal use of steganography, steganalysis technology came into being. As a countermeasure technology for steganography, steganalysis analyzes the statistical characteristics of the carrier to determine whether there is additional information hidden in the carrier, and even estimates the amount of embedded secret information and obtains the content of the secret information. The research on steganalysis is of great significance in preventing the leakage of confidential information, combating terrorism and criminal activities, and maintaining Internet security, and has been highly valued by governments and scientific research institutions of various countries.

[0003] Since digital video data is generally large, it is usually transmitted in a compressed form in an open transmission channel. H.265 / HEVC (High Efficiency Video Coding) is the latest generation of video compression coding standards. Its core goal is to double the compression efficiency based on H.264 / AVC, and it plays an important role in high-definition, ultra-high-definition and other video application fields.

[0004] However, compared with steganalysis research in the fields of text, images, and audio, due to the complexity of H.265 / HEVC encoding, the current video steganalysis method for HEVC transform coefficients is still in its infancy. Existing research mainly relies on the correlation theory between pixels to decode the compressed video into the spatial domain, and extracts the statistical correlation features of the video in the spatial domain and the temporal domain by transferring the relevant methods in the field of image steganalysis. Due to the differences between compressed videos and images, this manual feature extraction method is often difficult to adapt to a variety of steganographic algorithms in actual application scenarios, and cannot effectively detect steganographic algorithms in low-embedded scenarios.

[0005] In recent years, the rapid development of deep learning technology has also provided new tools for steganalysis technology. By building more complex and powerful models, deep learning technology can improve the accuracy and robustness of steganalysis and effectively deal with complex network security threats. However, there is still a lack of HEVC transform coefficient video steganalysis combined with deep learning technology. Therefore, it is of great practical significance to study a method of steganalysis under the H.265 / HEVC coding standard combined with deep learning technology. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a HEVC transform coefficient video steganalysis method based on an attention mechanism, which combines deep learning technology for steganalysis, can improve the accuracy and robustness of steganalysis, can adapt to a variety of steganography methods, and has high detection performance in low embedding rate scenarios.

[0007] The technical solution adopted by the present invention to solve the above technical problems is: a HEVC transform coefficient video steganalysis method based on an attention mechanism, characterized by comprising the following steps:

[0008] Step 1: Obtain several compressed videos, wherein any compressed video is a compressed original video obtained by encoding with the H.265 / HEVC standard encoder or a compressed steganographic video obtained by encoding with the H.265 / HEVC standard encoder and then steganographically; then perform data preprocessing on each compressed video, the data preprocessing includes two stages: data selection and data enhancement. The data selection is to extract all I frames in the decoding process of each compressed video and obtain the prediction residual map corresponding to each I frame. The data enhancement is to randomly flip the prediction residual map corresponding to each I frame, wherein the size of the prediction residual map is H×W×1; then normalize the value range of the flipped prediction residual map corresponding to each I frame to [0,1]; finally, the prediction residual maps corresponding to the I frames extracted in the decoding process of all compressed videos and in the value range [0,1] constitute a training set;

[0009] Step 2: Construct a deep neural network as the transform coefficient video steganalysis network: The transform coefficient video steganalysis network consists of a convolutional sub-network for extracting steganalysis traces within video frames, a self-attention sub-network for strengthening steganalysis frames, and a binary classifier for determining whether the compressed video is a compressed original video or a compressed steganalysis video;

[0010] The convolutional subnetwork adopts an architecture based on the convolutional residual network, which is divided into two stages: feature extraction and feature representation. The feature extraction stage consists of a preprocessing block and three sequentially connected shallow feature extraction blocks, and the feature representation stage consists of two downsampling blocks based on the attention mechanism, two basic feature extraction blocks and a global average pooling. The prediction residual map corresponding to an I frame extracted during the decoding process of a compressed video has a value range of [0,1] and a size of H×W×1. After the preprocessing block, the residual feature map of the centered error with a size of H×W×16 is obtained. The residual feature map of the centered error is obtained after three sequentially connected shallow feature extraction blocks. A fused shallow feature map of size H×W×16 is obtained after the fused shallow feature map passes through the first downsampling block based on the attention mechanism to obtain a first attention feature map of size H×W×32. The first attention feature map passes through the first basic feature extraction block to obtain a first basic feature map of size H×W×32. The first basic feature map passes through the second downsampling block based on the attention mechanism to obtain a second attention feature map of size H×W×64. The second attention feature map passes through the second basic feature extraction block to obtain a second basic feature map of size H×W×64. The second basic feature map is subjected to global average pooling to obtain a feature vector of size 64×1.

[0011] Assume that N I frames are extracted during the decoding process of a compressed video. After the convolution sub-network outputs N feature vectors of size 64×1 corresponding to a compressed video, the self-attention sub-network receives N feature vectors of size 64×1 corresponding to a compressed video, and the self-attention sub-network outputs a video feature vector of size N×64×1.

[0012] The predicted residual graphs corresponding to the N I frames extracted during the decoding process of a compressed video with a value range of [0,1] are passed through the convolution sub-network and the self-attention sub-network to obtain a video feature vector of size N×64×1. The result of the splicing operation on the channel dimension is input into the binary classifier, which determines whether the compressed video is a compressed original video or a compressed steganographic video;

[0013] Step 3: Use the training set to train the transformation coefficient video steganalysis network, and obtain the trained transformation coefficient video steganalysis network after the training is completed;

[0014] Step 4: For any test compressed video, all I frames are extracted during the decoding process of the test compressed video, and the prediction residual map corresponding to each I frame is obtained, wherein the test compressed video is a compressed original video encoded by the H.265 / HEVC standard encoder or a compressed steganographic video encoded by the H.265 / HEVC standard encoder and then steganographically obtained; then the prediction residual maps corresponding to all I frames extracted during the decoding process of the test compressed video are input into the trained transform coefficient video steganalysis network, and the trained transform coefficient video steganalysis network outputs a determination result of whether the test compressed video is a compressed original video or a compressed steganographic video.

[0015] In step 1, the specific process of data selection is: use the H.265 / HEVC standard decoder to decode each compressed video, extract all I frames during the decoding process, and save the prediction residual value corresponding to each pixel in each I frame to obtain the prediction residual map corresponding to each I frame; flipping is vertical flipping or horizontal flipping.

[0016] In the step 2, the implementation process of the preprocessing block is as follows: the prediction residual map with a value range of [0,1] and a size of H×W×1 is extracted after passing through a learnable SRM high-pass filter kernel to obtain the residual connection of the centering error, wherein the convolution kernel size of the convolution in the learnable SRM high-pass filter kernel is 5×5; the residual connection of the centering error is successively subjected to a batch normalization, a truncation function TLU and a convolution kernel size of 1×1 to obtain a residual feature map of the centering error with a size of H×W×16.

[0017] In the step 2, the structures of the three shallow feature extraction blocks are the same, and each shallow feature extraction block has two feature extraction branches and a feature fusion branch; the implementation process of the first feature extraction branch is: the input feature map is convolved with a convolution kernel size of 1×1 to obtain a first shallow feature map; the second feature extraction branch uses a convolutional residual network architecture, and the implementation process of the second feature extraction branch is: the input feature map is sequentially convolved with a first convolution kernel size of 3×3, the first batch normalization, the Relu activation function, the second convolution kernel size of 3×3, and the second batch normalization to obtain an intermediate feature map, and the intermediate feature map is then added to the input feature map to obtain a second shallow feature map; the implementation process of the feature fusion branch is: The first shallow feature map and the second shallow feature map are concatenated in the channel dimension and then subjected to a convolution with a convolution kernel size of 1×1 to obtain a fused shallow feature map; wherein, the sizes of the input feature map, the first shallow feature map, the intermediate feature map, the second shallow feature map, and the fused shallow feature map are all H×W×16, the input feature map in the first shallow feature extraction block is the residual feature map of the centralization error, the input feature map in the second shallow feature extraction block is the fused shallow feature map output by the first shallow feature extraction block, the input feature map in the third shallow feature extraction block is the fused shallow feature map output by the second shallow feature extraction block, and the fused shallow feature map output by the third shallow feature extraction block is used to input the first downsampling block based on the attention mechanism.

[0018] In the step 2, the structures of the two downsampling blocks based on the attention mechanism are the same, and each downsampling block based on the attention mechanism is divided into two feature representation branches; the implementation process of the first feature representation branch is: the input feature map is sequentially subjected to a convolution with a convolution kernel size of 1×1 and a batch normalization to obtain a first feature map; the implementation process of the second feature representation branch is: the input feature map is sequentially subjected to a convolution with a convolution kernel size of 3×3, the first batch normalization, the Relu activation function, an average pooling, the second batch normalization, and a central error perception block to obtain a second feature map; the first feature map and the second feature map are subjected to The output feature map is obtained by the addition operation; wherein, in the first downsampling block based on the attention mechanism, the input feature map is the fused shallow feature map of size H×W×16 output by the third shallow feature extraction block, the sizes of the first feature map, the second feature map, and the output feature map are all H×W×32, and the output feature map is used as the first attention feature map; in the second downsampling block based on the attention mechanism, the input feature map is the first basic feature map of size H×W×32 output by the first basic feature extraction block, the sizes of the first feature map, the second feature map, and the output feature map are all H×W×64, and the output feature map is used as the second attention feature map.

[0019] The central error perception block is divided into two stages: channel attention and spatial attention;

[0020] In the channel attention stage, the input feature map is sequentially subjected to a global average pooling, a multi-layer perceptron and a softmax function to obtain the channel weight, which is then multiplied by the input feature map to obtain the channel attention feature map. In the centralized error perception block in the first downsampling block based on the attention mechanism, the size of the channel weight is 1×1×16, and the size of the channel attention feature map is H×W×16; in the centralized error perception block in the second downsampling block based on the attention mechanism, the size of the channel weight is 1×1×32, and the size of the channel attention feature map is H×W×32;

[0021] In the spatial attention stage, the channel attention feature map is successively subjected to a convolution with a convolution kernel size of 1×1 and a softmax function to obtain a spatial attention weight map, and the spatial attention weight map is then multiplied with the channel attention feature map to obtain a weighted feature map. In the centralized error perception block in the first downsampling block based on the attention mechanism, the size of the spatial attention weight map is H×W×1, and the weighted feature map is used as the second feature map in the first downsampling block based on the attention mechanism, with a size of H×W×32; in the centralized error perception block in the second downsampling block based on the attention mechanism, the size of the spatial attention weight map is H×W×1, and the weighted feature map is used as the second feature map in the second downsampling block based on the attention mechanism, with a size of H×W×64.

[0022] In the step 2, the structures of the two basic feature extraction blocks are the same, and the implementation process of each basic feature extraction block is as follows: the input feature map is sequentially subjected to the first convolution with a convolution kernel size of 3×3, the first batch normalization, the Relu activation function, the second convolution with a convolution kernel size of 3×3, and the second batch normalization to obtain an intermediate feature map, and the intermediate feature map is then added to the input feature map to obtain an output feature map; in the first basic feature extraction block, the input feature map is a first attention feature map with a size of H×W×32 output by the first downsampling block based on the attention mechanism, the sizes of the intermediate feature map and the output feature map are both H×W×32, and the output feature map is used as the first basic feature map; in the second basic feature extraction block, the input feature map is a second attention feature map with a size of H×W×64 output by the second downsampling block based on the attention mechanism, the sizes of the intermediate feature map and the output feature map are both H×W×64, and the output feature map is used as the second basic feature map.

[0023] In step 3, cross entropy loss is used during training and the gradient is updated.

[0024] Compared with the prior art, the advantages of the present invention are:

[0025] 1) The method of the present invention takes the prediction residual map as the analysis object and proposes a transformation coefficient video steganalysis network based on the attention mechanism, which solves the problem of low detection performance of the current related transformation coefficient steganalysis.

[0026] 2) To address the problem of sparse steganographic traces within and between frames caused by adaptive steganography, the method of the present invention proposes a centralized error perception block and a self-attention mechanism to adaptively enhance the steganographic trace area and steganographic frame features. Therefore, the method of the present invention has higher robustness and accuracy in detection scenarios of various steganographic methods and low embedding rates. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flowchart of the overall implementation process of the method of the present invention;

[0028] Figure 2 Schematic diagram of the composition structure of the convolutional sub-network;

[0029] Figure 3 Schematic diagram of the composition structure of the centralized error perception block. DETAILED DESCRIPTION

[0030] The present invention is further described in detail below with reference to the accompanying drawings.

[0031] The present invention proposes a HEVC transform coefficient video steganalysis method based on the attention mechanism, and its overall implementation flow chart is as follows: Figure 1 As shown, it includes the following steps:

[0032] Step 1: Obtain several compressed videos, wherein any compressed video is a compressed original video obtained by encoding with the H.265 / HEVC standard encoder or a compressed steganographic video obtained by encoding with the H.265 / HEVC standard encoder and then steganographically; then perform data preprocessing on each compressed video, the data preprocessing includes two stages: data selection and data enhancement. The data selection is to extract all I frames in the decoding process of each compressed video and obtain the prediction residual map corresponding to each I frame. The data enhancement is to randomly flip the prediction residual map corresponding to each I frame, wherein the size of the prediction residual map is H×W×1; then normalize the value range of the flipped prediction residual map corresponding to each I frame to [0,1]; finally, the prediction residual maps corresponding to the I frames extracted in the decoding process of all compressed videos and whose value range is in [0,1] constitute a training set.

[0033] In the present invention, the specific process of data selection is: use the H.265 / HEVC standard decoder to decode each compressed video, extract all I frames during the decoding process, and save the prediction residual value corresponding to each pixel point in each I frame to obtain the prediction residual map corresponding to each I frame; flipping is vertical flipping or horizontal flipping.

[0034] Step 2: Construct a deep neural network as the transformation coefficient video steganalysis network: Figure 1 As shown in the figure, the transform coefficient video steganalysis network consists of a convolutional subnetwork (CESNet) for extracting steganalysis traces in video frames, a self-attention subnetwork (Self Attention) for strengthening steganalysis frames, and a binary classifier (Classifier) ​​for determining whether the compressed video is a compressed original video (Cover) or a compressed stego video (Stego).

[0035] like Figure 2 As shown in the figure, the convolutional subnetwork adopts an architecture based on a convolutional residual network, which is divided into two stages: feature extraction and feature representation. The feature extraction stage consists of a preprocessing block (PreprocessBlock) and three shallow feature extraction blocks (ShallowBlock) connected in sequence. The feature representation stage consists of two downsampling blocks (AttentionBlock) based on the attention mechanism, two basic feature extraction blocks (BasicBock) and a global average pooling. The prediction residual map (Prediction Residual Map) corresponding to an I frame extracted during the decoding process of a compressed video has a value range of [0,1] and a size of H×W×1. Map) obtains a residual feature map of the centering error with a size of H×W×16 after the preprocessing block, and the residual feature map of the centering error obtains a fused shallow feature map with a size of H×W×16 after three sequentially connected shallow feature extraction blocks. The fused shallow feature map obtains a first attention feature map with a size of H×W×32 after the first downsampling block based on the attention mechanism, and obtains a first basic feature map with a size of H×W×32 after the first attention feature map passes through the first basic feature extraction block. The first basic feature map obtains a second attention feature map with a size of H×W×64 after the second downsampling block based on the attention mechanism, and the second attention feature map obtains a second basic feature map with a size of H×W×64 after the second basic feature extraction block. The second basic feature map obtains a feature vector with a size of 64×1 after global average pooling.

[0036] Further limiting, such as Figure 2As shown in FIG. 1 , the implementation process of the preprocessing block is as follows: the prediction residual map with a value range of [0,1] and a size of H×W×1 is extracted after passing through a learnable SRM high-pass filter kernel (Conv 5×5) to obtain the residual connection of the centering error, wherein the convolution kernel size of the convolution in the learnable SRM high-pass filter kernel is 5×5; the residual connection of the centering error is sequentially subjected to a batch normalization (Norm), a truncation function TLU and a convolution kernel size of 1×1 (Conv 1×1) to obtain a residual feature map of the centering error with a size of H×W×16. Here, the learnable SRM high-pass filter kernel is an existing technology, which includes 30 convolutions with a convolution kernel size of 5×5, a total of first-order to third-order residual filter kernels, and has rich residual feature representations, so that the network can be in a good initial state.

[0037] Further limiting, such as Figure 2 As shown in the figure, the structures of the three shallow feature extraction blocks are the same. Each shallow feature extraction block has two feature extraction branches and one feature fusion branch. The implementation process of the first feature extraction branch is as follows: the input feature map is subjected to a convolution with a convolution kernel size of 1×1 (Conv 1×1) to obtain the first shallow feature map. The second feature extraction branch uses the classic convolutional residual network architecture. The implementation process of the second feature extraction branch is as follows: the input feature map is subjected to the first convolution with a convolution kernel size of 3×3 (Conv 3×3), the first batch normalization (Norm), the Relu activation function, the second convolution with a convolution kernel size of 3×3 (Conv 3×3), the second batch normalization (Norm) is performed to obtain the intermediate feature map, and the intermediate feature map is then added to the input feature map to obtain the second shallow feature map; the implementation process of the feature fusion branch is as follows: the first shallow feature map and the second shallow feature map are concatenated (concat operation) in the channel dimension and then passed through a convolution with a convolution kernel size of 1×1 (Conv 1×1) to obtain a fused shallow feature map; wherein, the size of the input feature map, the first shallow feature map, the intermediate feature map, the second shallow feature map, and the fused shallow feature map are all H×W×16, the input feature map in the first shallow feature extraction block is the residual feature map of the centralization error, the input feature map in the second shallow feature extraction block is the fused shallow feature map output by the first shallow feature extraction block, the input feature map in the third shallow feature extraction block is the fused shallow feature map output by the second shallow feature extraction block, and the fused shallow feature map output by the third shallow feature extraction block is used to input the first downsampling block based on the attention mechanism.

[0038] Further limiting, such as Figure 2As shown in the figure, the structures of the two downsampling blocks based on the attention mechanism are the same. Each downsampling block based on the attention mechanism is divided into two feature representation branches. The implementation process of the first feature representation branch is as follows: the input feature map is sequentially subjected to a convolution with a convolution kernel size of 1×1 (Conv 1×1) and a batch normalization (Norm) to obtain the first feature map; the implementation process of the second feature representation branch is as follows: the input feature map is sequentially subjected to a convolution with a convolution kernel size of 3×3 (Conv 3×3), the first batch normalization (Norm), Relu activation function, an average pooling (AvgPool), the second batch normalization (Norm), and a central error perception block (CEA) to obtain the second feature map; the first feature map and the second feature map are added to obtain the output feature map; wherein, in the first downsampling block based on the attention mechanism, the input feature map is the fused shallow feature map of size H×W×16 output by the third shallow feature extraction block, the sizes of the first feature map, the second feature map, and the output feature map are all H×W×32, and the output feature map is used as the first attention feature map; in the second downsampling block based on the attention mechanism, the input feature map is the first basic feature map of size H×W×32 output by the first basic feature extraction block, the sizes of the first feature map, the second feature map, and the output feature map are all H×W×64, and the output feature map is used as the second attention feature map.

[0039] like Figure 3As shown in the figure, the above-mentioned centralized error perception block is divided into two stages: channel attention and spatial attention. In the channel attention stage, the input feature map is sequentially subjected to a global average pooling (GAP), a multi-layer perceptron (MLP) and a softmax function to obtain the channel weight, which is then multiplied with the input feature map to strengthen the effective centralized error feature channel to obtain the channel attention feature map. Here, the input feature map channel is subjected to a global average pooling operation to obtain the importance of each channel, and the weight of each channel is calculated using a shared multi-layer perceptron. In the centralized error perception block in the first downsampling block based on the attention mechanism, the size of the channel weight is 1×1×16, and the size of the channel attention feature map is H×W×16; in the centralized error perception block in the second downsampling block based on the attention mechanism, the size of the channel weight is 1×1×32, and the size of the channel attention feature map is H×W×32. In the spatial attention stage, the channel attention feature map is successively subjected to a convolution with a convolution kernel size of 1×1 (Conv 1×1) and a softmax function to obtain a spatial attention weight map, and the spatial attention weight map is then multiplied with the channel attention feature map to obtain a weighted feature map. Here, a convolution with a convolution kernel size of 1×1 is used to obtain the importance of each spatial position, and the softmax function is used to obtain the corresponding spatial attention weight map. In the centralized error perception block in the first downsampling block based on the attention mechanism, the size of the spatial attention weight map is H×W×1, and the weighted feature map is used as the second feature map in the first downsampling block based on the attention mechanism, with a size of H×W×32; in the centralized error perception block in the second downsampling block based on the attention mechanism, the size of the spatial attention weight map is H×W×1, and the weighted feature map is used as the second feature map in the second downsampling block based on the attention mechanism, with a size of H×W×64. Here, a centralized error perception block is designed to solve the sparsity of steganographic traces in video frames; the multi-layer perceptron is an existing technology and has a two-layer structure by default; through the fusion of channel attention and spatial attention, the centralized error perception block can further extract the steganographic features of the centralized error.

[0040] It is further defined that the structures of the two basic feature extraction blocks are the same, and the implementation process of each basic feature extraction block is: the input feature map is sequentially subjected to the first convolution with a kernel size of 3×3 (Conv 3×3), the first batch normalization (Norm), the Relu activation function, the second convolution with a kernel size of 3×3 (Conv 3×3), and the second batch normalization (Norm) to obtain an intermediate feature map, and the intermediate feature map is then added to the input feature map to obtain an output feature map; in the first basic feature extraction block, the input feature map is the first attention feature map of size H×W×32 output by the first downsampling block based on the attention mechanism, the sizes of the intermediate feature map and the output feature map are both H×W×32, and the output feature map is used as the first basic feature map; in the second basic feature extraction block, the input feature map is the second attention feature map of size H×W×64 output by the second downsampling block based on the attention mechanism, the sizes of the intermediate feature map and the output feature map are both H×W×64, and the output feature map is used as the second basic feature map.

[0041] like Figure 1 As shown in the figure, assume that N I frames are extracted during the decoding process of a compressed video. After the convolution sub-network outputs N feature vectors of size 64×1 corresponding to a compressed video, the self-attention sub-network receives N feature vectors of size 64×1 corresponding to a compressed video, and the self-attention sub-network outputs a video feature vector of size N×64×1. Here, N represents the total number of I frames extracted during the decoding process of a compressed video, which is generally 4.

[0042] For the self-attention sub-network, the existing self-attention network structure is adopted. In the self-attention network structure, for the current I frame, the query vector Q and the value vector V are used to calculate the correlation weight between the feature vector of size 64×1 corresponding to the current I frame and the feature vector of size 64×1 corresponding to the remaining I frames, and then the calculated correlation weight is weighted and summed with the key vector K to obtain the video feature vector after self-attention enhancement.

[0043] like Figure 1 As shown in the figure, the predicted residual graph corresponding to the N I frames extracted during the decoding process of a compressed video with a value range of [0,1] is passed through the convolution sub-network and the self-attention sub-network to obtain a video feature vector of size N×64×1. The result of the splicing operation on the channel dimension is input into the binary classifier, which determines whether the compressed video is a compressed original video or a compressed steganographic video.

[0044] Step 3: Use the training set to train the transform coefficient video steganalysis network, and obtain the trained transform coefficient video steganalysis network after the training. The cross entropy loss is used in the training process, and the gradient is updated.

[0045] Step 4: For any test compressed video, all I frames are extracted during the decoding process of the test compressed video, and the prediction residual map corresponding to each I frame is obtained, wherein the test compressed video is a compressed original video encoded by the H.265 / HEVC standard encoder or a compressed steganographic video encoded by the H.265 / HEVC standard encoder and then steganographically obtained; then the prediction residual maps corresponding to all I frames extracted during the decoding process of the test compressed video are input into the trained transform coefficient video steganalysis network, and the trained transform coefficient video steganalysis network outputs a determination result of whether the test compressed video is a compressed original video or a compressed steganographic video.

[0046] In order to further illustrate the feasibility and effectiveness of the method of the present invention, a simulation experiment was carried out on the method of the present invention.

[0047] In the simulation experiment, the video sequences used are all from the Internet https: / / media.xiph.org / video / derf / , including 116 4:2:0 YUV video sequences with a resolution of 288×352, and each YUV video sequence contains 261 frames on average. The original YUV video sequence is encoded using the HEVC reference software HM 16.15, where the GOP is the IPPP structure and the encoding configuration file is encoder_lowdelay_P_main.cfg.

[0048] In the simulation experiment, three different steganography methods are used to hide the encoded video sequence to obtain compressed steganographic videos, and then three different steganalysis methods are used to detect the compressed steganographic videos. The three different steganography methods are the steganography method (referred to as Tar1) from reference 1 (Y.Liu, S.Liu, H.Zhao, and S.Liu, "A new data hiding method for H.265 / HEVC video streams without intra-frame distortion drift," Multimedia Tools and Applications, vol. 78, no. 6, pp. 6459-6486, 2019.), the steganography method (referred to as Tar1) from reference 2 (L.Yang, D.Xu, R.Wang, and S.He, "Adaptive HEVC video steganography based on distortion compensation optimization," Journal of Information Security andApplications,vol.73,2023.(Adaptive HEVC video steganography method based on distortion compensation optimization, Journal of Information Security and Applications)) (referred to as Tar2), the steganography method from document 3 (L.Yang,R.Wang,D.Xu,L.Dong andS.He,“Centralized Error Distribution-Preserving Adaptive Steganography forHEVC,”IEEE Transactions on Multimedia,vol.26,pp.4255-4270,2024(Adaptive steganography method based on centralized error preservation, IEEE Journal of Multimedia)) (referred to as Tar3).The three different steganalysis methods are the steganalysis method based on the centralized error under the HEVC platform from document 2 (abbreviated as MCER), the spatiotemporal feature steganalysis method based on accordion unfolding (AU) and special frames extraction (SFE) from document 4 (H. Shi, T. Sun, X. Jiang, et al, "AHEVC Video Steganalysis Against DCT / DST-Based Steganography," International Journal of Digital Crime and Forensics, vol. 13, no. 3, pp. 19-33, 2021. (A HEVC video steganalysis method for transform coefficients, International Journal of Digital Crime and Forensics)) (abbreviated as Shi), and the method of the present invention.

[0049] In the simulation experiment, the accuracy Acc is used as the evaluation index. The higher the Acc value, the stronger the detection performance.

[0050] Table 1 shows the comparison of the accuracy rates Acc of detecting the steganographic compressed videos obtained by Tar1, Tar2 and Tar3 steganography using MCER, Shi and the method of the present invention respectively.

[0051] Table 1 Accuracy of detecting the steganographic compressed videos obtained by Tar1, Tar2 and Tar3 steganography using MCER, Shi and the method of the present invention respectively Acc

[0052]

[0053] As can be seen from Table 1, the method of the present invention can not only effectively detect the steganographic methods of Tar1, Tar2, and Tar3, but also the detection performance on the three steganographic methods far exceeds that of other comparison methods. Therefore, the present invention has higher performance for compressed steganographic videos and has more robust characteristics in low embedding rate scenarios.

Claims

1. A HEVC transform coefficient video steganalysis method based on attention mechanism, characterized in that The following steps are involved: Step 1: Obtain several compressed videos, wherein any compressed video is a compressed original video obtained by encoding with the H.265 / HEVC standard encoder or a compressed steganographic video obtained by encoding with the H.265 / HEVC standard encoder and then steganographically; then perform data preprocessing on each compressed video, the data preprocessing includes two stages: data selection and data enhancement. The data selection is to extract all I frames in the decoding process of each compressed video and obtain the prediction residual map corresponding to each I frame. The data enhancement is to randomly flip the prediction residual map corresponding to each I frame, wherein the size of the prediction residual map is H×W×1; then normalize the value range of the flipped prediction residual map corresponding to each I frame to [0,1]; finally, the prediction residual maps corresponding to the I frames extracted in the decoding process of all compressed videos and in the value range [0,1] constitute a training set; Step 2: Construct a deep neural network as the transform coefficient video steganalysis network: The transform coefficient video steganalysis network consists of a convolutional sub-network for extracting steganalysis traces within video frames, a self-attention sub-network for strengthening steganalysis frames, and a binary classifier for determining whether the compressed video is a compressed original video or a compressed steganalysis video; The convolutional subnetwork adopts an architecture based on the convolutional residual network, which is divided into two stages: feature extraction and feature representation. The feature extraction stage consists of a preprocessing block and three sequentially connected shallow feature extraction blocks, and the feature representation stage consists of two downsampling blocks based on the attention mechanism, two basic feature extraction blocks and a global average pooling. The prediction residual map corresponding to an I frame extracted during the decoding process of a compressed video has a value range of [0,1] and a size of H×W×1. After the preprocessing block, the residual feature map of the centered error with a size of H×W×16 is obtained. The residual feature map of the centered error is obtained after three sequentially connected shallow feature extraction blocks. A fused shallow feature map of size H×W×16 is obtained after the fused shallow feature map passes through the first downsampling block based on the attention mechanism to obtain a first attention feature map of size W×W×32. The first attention feature map passes through the first basic feature extraction block to obtain a first basic feature map of size H×W×32. The first basic feature map passes through the second downsampling block based on the attention mechanism to obtain a second attention feature map of size H×W×64. The second attention feature map passes through the second basic feature extraction block to obtain a second basic feature map of size H×W×64. The second basic feature map is subjected to global average pooling to obtain a feature vector of size 64×1. Assume that N I frames are extracted during the decoding process of a compressed video. After the convolution sub-network outputs N feature vectors of size 64×1 corresponding to a compressed video, the self-attention sub-network receives N feature vectors of size 64×1 corresponding to a compressed video, and the self-attention sub-network outputs a video feature vector of size N×64×1. The predicted residual graphs corresponding to the N I frames extracted during the decoding process of a compressed video with a value range of [0,1] are passed through the convolution sub-network and the self-attention sub-network to obtain a video feature vector of size N×64×1. The result of the splicing operation on the channel dimension is input into the binary classifier, which determines whether the compressed video is a compressed original video or a compressed steganographic video; Step 3: Use the training set to train the transformation coefficient video steganalysis network, and obtain the trained transformation coefficient video steganalysis network after the training is completed; Step 4: For any test compressed video, all I frames are extracted during the decoding process of the test compressed video, and the prediction residual map corresponding to each I frame is obtained, wherein the test compressed video is a compressed original video encoded by the H.265 / HEVC standard encoder or a compressed steganographic video encoded by the H.265 / HEVC standard encoder and then steganographically obtained; then the prediction residual maps corresponding to all I frames extracted during the decoding process of the test compressed video are input into the trained transform coefficient video steganalysis network, and the trained transform coefficient video steganalysis network outputs a determination result of whether the test compressed video is a compressed original video or a compressed steganographic video.

2. The HEVC transform coefficient video steganalysis method based on the attention mechanism according to claim 1 is characterized in that In step 1, the specific process of data selection is: use the H.265 / HEVC standard decoder to decode each compressed video, extract all I frames during the decoding process, and save the prediction residual value corresponding to each pixel in each I frame to obtain the prediction residual map corresponding to each I frame; flipping is vertical flipping or horizontal flipping.

3. The HEVC transform coefficient video steganalysis method based on the attention mechanism according to claim 1 is characterized in that In the step 2, the implementation process of the preprocessing block is as follows: the prediction residual map with a value range of [0,1] and a size of H×W×1 is extracted after passing through a learnable SRM high-pass filter kernel to obtain the residual connection of the centering error, wherein the convolution kernel size of the convolution in the learnable SRM high-pass filter kernel is 5×5; the residual connection of the centering error is successively subjected to a batch normalization, a truncation function TLU and a convolution kernel size of 1×1 to obtain a residual feature map of the centering error with a size of H×W×16.

4. The HEVC transform coefficient video steganalysis method based on the attention mechanism according to claim 1 is characterized in that In the step 2, the structures of the three shallow feature extraction blocks are the same, and each shallow feature extraction block has two feature extraction branches and a feature fusion branch; the implementation process of the first feature extraction branch is: the input feature map is convolved with a convolution kernel size of 1×1 to obtain a first shallow feature map; the second feature extraction branch uses a convolutional residual network architecture, and the implementation process of the second feature extraction branch is: the input feature map is sequentially convolved with a first convolution kernel size of 3×3, the first batch normalization, the Relu activation function, the second convolution kernel size of 3×3, and the second batch normalization to obtain an intermediate feature map, and the intermediate feature map is then added to the input feature map to obtain a second shallow feature map; the implementation process of the feature fusion branch is: The first shallow feature map and the second shallow feature map are concatenated in the channel dimension and then subjected to a convolution with a convolution kernel size of 1×1 to obtain a fused shallow feature map; wherein, the sizes of the input feature map, the first shallow feature map, the intermediate feature map, the second shallow feature map, and the fused shallow feature map are all H×W×16, the input feature map in the first shallow feature extraction block is the residual feature map of the centralization error, the input feature map in the second shallow feature extraction block is the fused shallow feature map output by the first shallow feature extraction block, the input feature map in the third shallow feature extraction block is the fused shallow feature map output by the second shallow feature extraction block, and the fused shallow feature map output by the third shallow feature extraction block is used to input the first downsampling block based on the attention mechanism.

5. The HEVC transform coefficient video steganalysis method based on the attention mechanism according to claim 1 is characterized in that In the step 2, the structures of the two downsampling blocks based on the attention mechanism are the same, and each downsampling block based on the attention mechanism is divided into two feature representation branches; The implementation process of the first feature representation branch is as follows: the input feature map is sequentially subjected to a convolution with a convolution kernel size of 1×1 and a batch normalization to obtain the first feature map; the implementation process of the second feature representation branch is as follows: the input feature map is sequentially subjected to a convolution with a convolution kernel size of 3×3, the first batch normalization, the Relu activation function, an average pooling, the second batch normalization, and a centralized error perception block to obtain the second feature map; the first feature map and the second feature map are added to obtain the output feature map; wherein, in the first attention mechanism based In the downsampling block, the input feature map is the fused shallow feature map of size H×W×16 output by the third shallow feature extraction block, the sizes of the first feature map, the second feature map, and the output feature map are all H×W×32, and the output feature map is used as the first attention feature map; in the second downsampling block based on the attention mechanism, the input feature map is the first basic feature map of size H×W×32 output by the first basic feature extraction block, the sizes of the first feature map, the second feature map, and the output feature map are all H×W×64, and the output feature map is used as the second attention feature map.

6. The HEVC transform coefficient video steganalysis method based on the attention mechanism according to claim 5 is characterized in that The central error perception block is divided into two stages: channel attention and spatial attention; In the channel attention stage, the input feature map is sequentially subjected to a global average pooling, a multi-layer perceptron and a softmax function to obtain the channel weight, which is then multiplied by the input feature map to obtain the channel attention feature map. In the centralized error perception block in the first downsampling block based on the attention mechanism, the size of the channel weight is 1×1×16, and the size of the channel attention feature map is H×W×16; in the centralized error perception block in the second downsampling block based on the attention mechanism, the size of the channel weight is 1×1×32, and the size of the channel attention feature map is H×W×32; In the spatial attention stage, the channel attention feature map is successively subjected to a convolution with a convolution kernel size of 1×1 and a softmax function to obtain a spatial attention weight map, and the spatial attention weight map is then multiplied with the channel attention feature map to obtain a weighted feature map. In the centralized error perception block in the first downsampling block based on the attention mechanism, the size of the spatial attention weight map is H×W×1, and the weighted feature map is used as the second feature map in the first downsampling block based on the attention mechanism, with a size of H×W×32; in the centralized error perception block in the second downsampling block based on the attention mechanism, the size of the spatial attention weight map is H×W×1, and the weighted feature map is used as the second feature map in the second downsampling block based on the attention mechanism, with a size of H×W×64.

7. The HEVC transform coefficient video steganalysis method based on the attention mechanism according to claim 1 is characterized in that In the step 2, the structures of the two basic feature extraction blocks are the same, and the implementation process of each basic feature extraction block is as follows: the input feature map is sequentially subjected to the first convolution with a convolution kernel size of 3×3, the first batch normalization, the Relu activation function, the second convolution with a convolution kernel size of 3×3, and the second batch normalization to obtain an intermediate feature map, and the intermediate feature map is then added to the input feature map to obtain an output feature map; in the first basic feature extraction block, the input feature map is a first attention feature map with a size of H×W×32 output by the first downsampling block based on the attention mechanism, the sizes of the intermediate feature map and the output feature map are both H×W×32, and the output feature map is used as the first basic feature map; in the second basic feature extraction block, the input feature map is a second attention feature map with a size of H×W×64 output by the second downsampling block based on the attention mechanism, the sizes of the intermediate feature map and the output feature map are both H×W×64, and the output feature map is used as the second basic feature map.

8. The HEVC transform coefficient video steganalysis method based on the attention mechanism according to claim 1 is characterized in that In step 3, cross entropy loss is used during training and the gradient is updated.

Images