CN105915916B - Video Steganalysis Method Based on Motion Vector Rate-distortion Performance Estimation - Google Patents

Abstract

The present invention relates to a kind of video steganalysis methods based on the estimation of motion vector distortion performance.Several frame groups will be divided into measured compressed video first, each frame group is made of continuous video frame, and any video frame belongs to and only belongs to some frame group.Then steganalysis feature extraction is carried out for the frame group that some includes several motion vectors, it include: for each motion vector in some frame group, obtain the set being made of the motion vector and its nearby motion vectors, the distortion performance of each motion vector in the set is calculated, preset steganalysis feature then is extracted to the frame group；Above-mentioned steps are repeated, steganalysis feature extraction successively is carried out to all frame groups of video to be measured.Then the classifier based on steganalysis feature is used, steganography classification judgement is carried out to each frame group in video to be measured.The present invention can effectively detect current existing motion vector field video steganography method.

Description

Video Steganalysis Method Based on Motion Vector Rate-distortion Performance Estimation

technical field

The invention relates to a video steganalysis method, in particular to a video steganalysis method based on motion vector rate-distortion (Rate-Distortion) performance estimation and its application in digital media security protection. It belongs to the information hiding sub-field in the field of information security technology.

Background technique

Modern information hiding technologies mainly include steganography, steganalysis and digital watermarking. Steganography mainly studies how to embed secret information into digital images, videos, audios and other digital multimedia files to achieve the purpose of covert communication; steganalysis mainly uses machine learning, pattern recognition and other methods to make steganographic classification judgments on the files to be tested.

With the widespread popularity of video-on-demand, Internet streaming, and handheld portable camera devices, digital video has become the most influential medium in the entertainment industry today. In addition, with the rapid development of technologies such as video compression, computer networks and high-performance computing, original video materials can be quickly compressed while ensuring high coding efficiency and visual fidelity, and transmitted in real time on the Internet. Therefore, digital video is an ideal carrier for steganographic communication.

Video steganography can be generally divided into Spatial Domain video steganography and Compressed Domain video steganography: the former embeds secret information by directly modifying the original pixel value of the video frame; the latter is used in the video compression coding process. Steganographic perturbation is introduced in , so that compression coding and steganographic embedding can be performed simultaneously. Compressed domain video steganography can be divided into Motion Vector-based video steganography, Transform Coefficient-based video steganography, and Intra Prediction Mode-based video according to different embedding domains. Steganography, video steganography based on Inter Prediction Mode, video steganography based on quantization process, and video steganography based on Entropy Coding. Among them, the steganographic embedding of secret information using the motion vector as a carrier has the following three advantages: First, the motion vector is generated in the motion estimation (Motion Estimation) module of the video, representing the original block and the corresponding prediction reference block (Prediction Reference Block) The disturbances introduced by the slight modification of the motion vector will be automatically processed by subsequent modules such as motion compensation, transform coding and entropy coding, which will only have an extremely small impact on the visual quality of the reconstructed video; secondly, Motion vectors in videos usually have a large range of values, which enables the design of motion vector domain video steganography methods that maintain the statistical properties of motion vectors; Vector video steganography can have a large steganographic embedding capacity.

Based on the above reasons, motion vector domain video steganography has been widely concerned by scholars in this field for a long time, and has experienced the following three development stages: the first generation of motion vector domain video steganography A portion of the motion vectors are used for steganographic embedding of secret information (Reference: F. Jordan, M. Kutter, and T. Ebrahimi, "Proposal of a watermarking technique for hiding/retrieving data in compressed and decompressed video," ISO/IEC Document, JTC1 /SC29/WG11, Stockholm, Sweden, Tech. Rep. M2281, Jul. 1997.). The second-generation motion vector domain video steganography method applies steganographic codes such as STC (Syndrome Trellis Codes, check grid code) and WPC (Wet Paper Codes, wet paper coding) to improve the embedding efficiency and steganographic security ( References: Y. Cao, X. Zhao, D. Feng, and R. Sheng, "Video steganography with perturbed motion estimation," in Proc.13thInt.Conf.Inf.Hiding,LNCS,vol.6958,Prague,Czech Republic , May 2011, pp.193–207.; Y. Yao, W. Zhang, N. Yu, and X. Zhao, “Defining embedding distortion for motionvector-based video steganography,” Multimedia Tools and Applications, vol.74, no. 24, pp.11 163–11 186, Dec. 2015.). However, since these methods inevitably destroy the local optimum of motion vectors in the process of steganographic embedding, they cannot resist the attack of AoSO, the most effective steganalysis method in the current motion vector domain (Reference: K. Wang, H. Zhao, and H. Wang, "Video steganalysis against motion vector-based steganography by adding or subtracting one motion vectorvalue," IEEE Trans.Inf.Forensics Security,vol.9,no.5,pp.741–751,May 2014 .).) The third-generation motion vector domain video steganography method, which is in the latest development stage, can try to maintain the local optimum of the modified motion vector during the embedding process of secret information. Therefore, compared with the previous two generations of methods, the anti-steganography method is greatly improved. Analysis performance (reference: Y.Cao, H.Zhang, X.Zhao, and H.Yu, "Video steganography based on optimizedmotion estimation perturbation," in Proc.3rd ACM Workshop Inf.HidingMultimedia Security,Portland,OR,USA, Jun. 2015, pp. 25–31.; H. Zhang, Y. Cao, and X. Zhao, “Motion vector-based video steganography with preserved localoptimality,” Multimedia Tools and Applications, 2015, Article in Press.).

After patent inquiry, the existing relevant patent applications in the field of the present invention are as follows:

(1) The Chinese patent with the patent application number of 2015103093163 "An Improved Video Steganalysis Method Based on Motion Vector Response" discloses a motion vector domain video steganalysis method. This method designs a steganalysis feature based on the fact that the motion vector in the steganographic video will recover after recompression, and obtains the key parameters of video compression and coding by reconstructing the first compression process, thereby effectively improving the practicability and stealth. Write analysis accuracy. Since the present invention does not involve recompressing the video, the design idea and specific implementation manner of this method are obviously different from those of the present invention.

(2) The Chinese patent "A Content-Adaptive Video Steganalysis Method" with the patent application number of 2015102222805 discloses a motion vector domain video steganalysis method. The method divides the video to be tested into variable-length detection intervals by calculating frame dynamics, so as to perform content-adaptive feature extraction and steganalysis. Since this method proposes a content-adaptive steganalysis feature extraction strategy and does not involve the design of specific steganalysis features, the basic purpose, design idea and specific implementation method of this method and the present invention are obviously different .

(3) The Chinese patent with the patent application number of 2013100660098 "Video steganalysis method based on local cost non-optimal statistics" discloses a motion vector domain video steganalysis method. The method determines whether the video has undergone steganography by calculating the variation of the local optimal probability of the motion vector before and after recompression of the video to be tested. Since the present invention does not involve recompressing the video, the design idea and specific implementation manner of this method are obviously different from those of the present invention.

SUMMARY OF THE INVENTION

The purpose of the present invention is to accurately determine whether the motion vector in the compressed video is locally optimal in the sense of rate distortion by accurately estimating the rate-distortion performance of the motion vector in the compressed video. Steganalysis performance of motion vector domain video steganalysis methods.

Compared with other motion vector domain video steganalysis methods, the present invention adopts the Lagrangian cost function to estimate the rate-distortion performance of motion vectors, and defines two different types of local optimal motion vectors in the sense of rate-distortion. A 36-dimensional motion vector domain steganalysis feature is designed. It can be seen that the method proposed by the present invention is different from any previous video steganalysis methods, and is especially suitable for digital multimedia security protection scenarios that require higher security levels.

According to the research, the existing motion vector domain video steganography analysis methods have the following two limitations: First, the design principles of most analysis methods do not make full use of the essential weakness of motion vector domain video steganography, that is, the The embedding modification of , will inevitably destroy its local optimum, thereby modifying the original locally optimum motion vector into a non-local optimum. Therefore, when the steganographic load rate is low, it is difficult for these methods to guarantee the ideal analysis and detection effect. Secondly, although some analysis methods are essentially correct, that is, the design of steganalysis features starts from detecting the local optimum of the motion vector, but they only judge whether the motion vector in the compressed video is the local optimum according to the distortion (Distortion), and The key role played by code rate estimation in the decision process is ignored. Since these methods fail to correctly detect the local optima of motion vectors in the rate-distortion sense, they cannot compete with the current state-of-the-art video steganography methods in the motion vector domain.

Specifically, the technical scheme adopted in the present invention is as follows:

A video steganalysis method based on motion vector rate-distortion performance estimation, comprising the following steps:

1) Frame group division: the compressed video to be tested is divided into several frame groups, each frame group is composed of continuous video frames, and any video frame belongs to and only belongs to a certain frame group;

2) for a certain frame group F _g that contains N motion vectors, perform steps 3) to 5) to perform steganalysis feature extraction;

3) Preprocessing: for each motion vector V _i =(x _i , y _i ) in the frame group F _g , where i=1, 2, _. The set composed of Ω(V _i );

4) Motion vector rate-distortion performance estimation: According to the preset motion vector rate-distortion performance estimation method, calculate the rate-distortion performance of each motion vector in the set Ω(V _i ), where i=1,2,...,N ;

5) feature calculation and extraction: according to the calculation result of step 4), the preset steganalysis feature is extracted to the frame group F _g ;

6) Repeat steps 2) to 5), successively carry out steganalysis feature extraction for all frame groups of the video to be tested;

7) Steganalysis: Using a classifier based on preset steganalysis features, steganographic classification judgment is performed on each frame group in the video to be tested.

On the basis of the above scheme, the present invention has further improved, that is, when the present invention is used for steganalysis (FIG. 1), for each motion vector in the compressed video to be tested V _i =(x _i ,y _i ) , the set Ω(V _i ) consisting of V _i and its adjacent motion vectors is obtained as follows:

Due to its simplicity and effectiveness, Lagrangian optimization techniques have been widely used in most video compression coding standards, including H.264/AVC and H.265/HEVC. The rate-distortion performance of video coding is efficiently optimized.

In the process of video compression and encoding, performing motion estimation (Motion Estimation) on a certain block (Block) refers to (Fig. 2): delineating a search range in the encoded reference frame (Reference Frame), The block is compared with all or part of the candidate blocks in the search range, and the best matching block is selected as the best prediction reference block of the original block to be coded. In particular, if a given Lagrangian multiplier λ is used to control the balance between distortion (Distortion) and code rate (Rate), it can be divided into two parts by minimizing the Lagrangian Cost Function (Lagrangian Cost Function) Block S performs motion estimation based on rate-distortion optimization, namely:

Among them, S' _m represents the prediction reference block corresponding to the sub-block S pointed to by the motion vector m, and D(S, S' _m ) represents the distortion between S and S' _m , which is generally calculated as the absolute error sum of the sub-block pixels. (Sum of Absolute Differences, SAD) measure, namely where k represents the position index of the pixel, A represents the set of position indices of all pixels in the block; R(m, ref_ind) represents the number of bits required to encode the motion vector m and the corresponding reference frame index ref_ind; Ω represents the motion in the reference frame Estimated search area. In addition to SAD, SATD (Sum of AbsoluteTransformed Differences) is also a commonly used block matching metric in motion estimation, namely Where T represents the Hadamard Transform, and α represents the normalization factor.

Based on the above introduction to the Lagrangian cost function and motion estimation in video compression coding, the motion vector rate-distortion performance estimation method proposed by the present invention, the definition of the local optimal motion vector in the sense of rate-distortion, and the 36-dimensional high-performance steganography The analysis features are described in detail below.

[1] Motion vector rate-distortion performance estimation method:

According to the motion estimation based on rate-distortion optimization in video coding, the present invention uses the Lagrangian cost function to estimate the rate-distortion performance of the motion vector in the compressed video.

Given a motion vector V, its rate-distortion performance can be estimated as follows:

J _D (V)=D(S _rec ,S _V )+λR(V).

Among them, S _rec represents the reconstructed block with the motion vector V (Reconstructed Block), S _V represents the prediction reference block corresponding to S _rec pointed to by the motion vector V, and D (S _rec , S _V ) represents the difference between S _rec and S _V Distortion between, λ is the Lagrangian multiplier, R(V) represents the number of bits required to encode the motion vector V.

[2] Definition of local optimal motion vector in the sense of rate distortion:

Based on the motion vector rate-distortion performance estimation, the present invention defines two locally optimal motion vectors in the sense of rate-distortion in compressed videos.

Definition 1: Type I local optimal motion vector. Given a motion vector V in the compressed video, if it satisfies

Then V is called the I-type local optimal motion vector.

Definition 2: Type II locally optimal motion vector. Given a motion vector V in the compressed video, if it satisfies

Then V is called a type II locally optimal motion vector.

In Definition 1 and Definition 2, Ω(V) represents the set consisting of V and its adjacent motion vectors, represents the reconstructed block with V, and S _m represents the motion vector m points to corresponding to The prediction reference block of λ is the Lagrange multiplier, and R(m) represents the number of bits required to encode the motion vector m.

[3] 36-dimensional high-performance motion vector domain steganalysis features:

Given a frame group F _g containing N motion vectors in the compressed video, four types of sub-features are extracted from F _g according to the following steps, and finally combined to obtain a 36-dimensional steganalysis feature.

a) Preprocessing: For each motion vector V _i =(x _i ,y _i )(i∈[1,N]) in F _g , first obtain the set of V _i and its adjacent motion vectors, namely Ω (V _i )={x _i -1,x _i ,x _i +1}×{y _i -1,y _i ,y _i +1}; and then for each motion vector in the set Ω(V _i ) where j∈[1,9] (Fig. 3), calculate its rate-distortion performance and Finally, determine the minimum value of the elements in the set {J _SAD (m)|m∈Ω(V _i )}, denoted as Similarly, get

b) Type 1 sub-feature extraction: each dimension of the type 1 sub-feature corresponds to the given k and equal probability, defined as

Among them, the function The same below.

c) Type 2 sub-feature extraction: Type 2 sub-features and J _SAD (V _i ) and The relative error between , is defined as

d) Type 3 sub-feature extraction: each dimension of the type 3 sub-feature corresponds to the given k and equal probability, defined as

e) Type 4 sub-feature extraction: Type 4 sub-feature with J _SATD (V _i ) and The relative error between , is defined as

f) Final feature merging: The final 36-dimensional steganalysis feature is obtained by merging the above 4 types of sub-features, and is defined as

Based on the above description, the video steganalysis method using the Lagrangian cost function to estimate the motion vector rate-distortion performance proposed by the present invention includes the following steps (if there is no special description, the following steps are all performed by a computer):

1) Frame group division: The compressed video to be tested is divided into several frame groups, each frame group is composed of consecutive video frames, and any video frame belongs to and only belongs to a certain frame group.

2) For a certain frame group F _g containing N motion vectors, perform steps 3) to 5) to perform steganalysis feature extraction.

3) Preprocessing: for each motion vector V _i =(x _i , y _i ) in the frame group F _g , where i=1, 2, _. The set composed of Ω(V _i )={x _i -1,x _i ,x _i +1}×{y _i -1,y _i ,y _i +1};

4) Motion vector rate-distortion performance estimation: for each motion vector in the set Ω(V _i ) where j∈[1,9] (Fig. 3), using SAD and SATD as the block matching metrics, respectively, the rate-distortion performance is calculated by the Lagrangian cost function and

5) Feature calculation and extraction: According to the calculation result of step 4), the preset 4 types of sub-features are extracted from the frame group F _g , and finally combined to obtain 36-dimensional steganalysis features.

6) Repeat steps 2) to 5) to sequentially perform steganalysis feature extraction for all frame groups of the video to be tested.

7) Steganalysis: Using a classifier based on preset steganalysis features, steganographic classification judgment is performed on each frame group in the video to be tested.

The beneficial effects of the video steganalysis method of the present invention to the relevant technical field are as follows:

1) It can effectively detect the existing motion vector domain video steganography methods. According to the rate-distortion optimization in video compression coding, any motion vector is locally optimal in the sense of rate-distortion, so any modification to the motion vector will inevitably destroy its local optimum in rate-distortion performance. Because the invention adopts the Lagrangian cost function to accurately estimate the rate-distortion performance of the motion vector in the compressed video, and designs the steganalysis feature on this basis, so as to make full use of the essence of the motion vector domain video steganography Weaknesses and grasp the key core of implementing motion vector domain video steganography analysis, so the present invention has ideal detection effect for all existing motion vector domain video steganography methods.

2) The carrier source mismatch phenomenon in steganalysis can be effectively alleviated to a certain extent. The phenomenon of Cover Source Mismatch in steganalysis refers to: when a steganalysis detector trained on a certain carrier source analyzes samples from different carrier sources, the difference between the two carrier sources is different. The difference will have a large negative impact on the accuracy of steganalysis. The carrier-source mismatch phenomenon is ubiquitous in the real network environment, which is the biggest obstacle to the practical application of steganalysis. Because the present invention makes full use of the essential weakness of video steganography in the motion vector domain, by accurately estimating the rate-distortion performance of the motion vector in the compressed video, the local optimal motion vector is accurately identified and the steganalysis performance and its stability are effectively improved, Therefore, the present invention can effectively alleviate the carrier source mismatch phenomenon to a certain extent. For example, the method of the present invention is used to extract steganalysis features for samples of the same size, the same code rate and the same embedding rate, and on this basis, the steganography obtained by training The analysis detector has good analysis and detection effect on steganographic videos of different sizes, different bit rates and different embedding rates.

3) Widely applicable to different video coding standards. Some motion vector domain video steganalysis methods are designed under the framework of a specific video coding standard and rely on the unique characteristics of the standard, which greatly affects the scope of application of these analysis methods. The present invention mainly recognizes the local optimal motion vector by accurately estimating the rate-distortion performance of the motion vector, thereby effectively improving the steganalysis performance, and its realization does not depend on a specific video coding standard, so the present invention has a larger scope of application, It can effectively analyze the motion vector domain video steganography based on different video coding standards.

4) It is scalable. When the present invention extracts steganalysis features from a frame group, for each motion vector V=(x, y), a set Ω(V)={x- 1,x,x+1}×{y-1,y,y+1}. Therefore, by modifying the calculation method of Ω(V), different video steganalysis methods based on motion vector rate-distortion performance estimation can be extended and customized to be applied to different steganalysis scenarios.

Description of drawings

Fig. 1 is the video steganalysis schematic diagram that adopts the present invention;

2 is a schematic diagram of block-based motion estimation in video compression coding;

3 is a schematic diagram of the relative spatial positions of all elements in the set Ω(V _i ) corresponding to a certain motion vector V _i ;

FIG. 4 is a flow chart of video steganalysis using the present invention.

Detailed ways

The present invention will be further described below through specific embodiments and in conjunction with FIG. 4 .

The video steganalysis method based on the motion vector rate-distortion performance estimation using the Lagrangian cost function proposed by the present invention, the specific operation details are as follows:

1) Frame group division: The compressed video to be tested is divided into several frame groups, each frame group is composed of consecutive video frames, and any video frame belongs to and only belongs to a certain frame group.

2) For a certain frame group F _g containing N motion vectors, perform steps 3) to 5) to perform steganalysis feature extraction.

3) Preprocessing: for each motion vector V _i =(x _i , y _i ) in the frame group F _g , where i=1, 2, _. The set composed of Ω(V _i )={x _i -1,x _i ,x _i +1}×{y _i -1,y _i ,y _i +1}, that is

4) Motion vector rate-distortion performance estimation: for each motion vector in the set Ω(V _i ) where j∈[1,9] (Fig. 3), using SAD and SATD as the block matching metrics, respectively, the rate-distortion performance is calculated by the Lagrangian cost function and which is

in, represents the reconstructed block with _Vi , represents the motion vector points to correspond to The prediction reference block of , λ is the Lagrange multiplier, represents the encoded motion vector required number of bits. Then, determine the minimum value of the elements in the set {J _SAD (m)|m∈Ω(V _i )}, denoted as Similarly, get

5) Feature calculation and extraction: According to the calculation result of step 4), using the feature extraction process described, extract the preset 4 types of sub-features for the frame group F _g so as to merge to obtain 36-dimensional steganalysis features.

6) Repeat steps 2) to 5) to sequentially perform steganalysis feature extraction for all frame groups of the video to be tested.

7) Steganalysis: Using a classifier based on the 36-dimensional steganalysis feature, steganographic classification judgment is performed on each frame group in the video to be tested.

It can be seen from the above specific embodiments: First, the present invention mainly identifies the local optimal motion vector in the compressed video by accurately estimating the rate-distortion performance of the motion vector and performs steganalysis, and its realization does not depend on specific video coding. Second, according to different practical application scenarios, different video steganalysis features based on motion vector rate-distortion performance estimation can be generated by changing the calculation method of Ω(·). Therefore, the present invention has a wider scope of application and greater flexibility.

In order to highlight that what the present invention proposes is a high-performance motion vector domain video steganalysis method, the following experimental configuration is used to conduct a steganalysis experiment:

1) YUV sequence: 250 YUV420 sequences with a resolution of CIF (352×288) and a length of 90-376 frames were collected through the Internet.

2) Video encoder and its configuration: The x264 open source video encoder is used to prepare compressed video samples. In order to reduce the time overhead, the encoding profile is set as the Baseline Profile.

3) Compressed video parameters: set the bitrate to 0.5mb/s, 3mb/s or 10mb/s, the frame rate is 50fps, and the motion estimation fast search algorithm is DIA, HEX or UMH.

4) Steganographic embedding rate: The Corrupted Motion Vector Ratio (CMVR) per frame is used to represent the steganographic embedding rate, and the CMVR of the carrier video is set to 0, and the CMVR of the steganographic video is set to 0.1 or 0.2.

5) Steganography method: The video steganography method proposed by Cao et al. in the motion vector domain with high concealment is selected for analysis (Reference: Y. "optimizedmotion estimation perturbation," in Proc. 3rd ACM Workshop Inf. Hiding Multimedia Security, Portland, OR, USA, Jun. 2015, pp. 25–31.).

6) Steganalysis method: Since AoSO is the most effective steganalysis method in the motion vector domain at present, it is compared with the present invention (reference: K. Wang, H. Zhao, and H. Wang, "Video steganalysis"). against motion vector-based steganography by adding or subtracting one motion vector value,” IEEE Trans. Inf. Forensics Security, vol. 9, no. 5, pp. 741–751, May 2014.).

7) Training and detection: In each group of steganalysis experiments, 60% of the "carrier-steganography" sample pairs were randomly selected for training the Support Vector Machine (SVM), and the remaining 40% of the sample pairs were used for steganography. To write classification decisions, each group of steganalysis experiments was repeated 50 times, and the obtained data were averaged.

According to the above experimental configuration, the obtained steganalysis results are shown in Table 1. It can be seen that the present invention can effectively detect the motion vector domain video steganography which is currently in the latest development stage and has the highest steganographic security. When the embedding rate and the code rate are both low, the steganalysis effect of the present invention is significantly better than that of AoSO, so the present invention is very suitable for the steganalysis scene with higher security level requirements.

Table 1. The average detection accuracy (%) of the method proposed by Cao et al. for steganalysis using AoSO and the present invention

The above embodiments are only used to illustrate the technical solutions of the present invention rather than limit them. Those of ordinary skill in the art can modify or equivalently replace the technical solutions of the present invention without departing from the spirit and scope of the present invention. The scope of protection shall be subject to what is stated in the claims.

Claims (4)

1. a video steganalysis method based on motion vector rate-distortion performance estimation, is characterized in that, comprises the following steps: 1) the compressed video to be tested is divided into several frame groups, each frame group is made up of continuous video frames, and any video frame belongs to and only belongs to a certain frame group; 2) for a certain frame group that contains several motion vectors, perform steps 3) to 5) to carry out steganalysis feature extraction; 3) For each motion vector V _i =(x _i ,y _i ) in a certain frame group F _g , where i=1,2,...,N, obtain the motion vector V _i and its adjacent A set of motion vectors Ω(V _i )={x _i -1,x _i ,x _i +1}×{y _i -1,y _i ,y _i +1}; 4) According to the preset motion vector rate-distortion performance estimation method, calculate the rate-distortion performance of each motion vector in the set, that is, for each motion vector in the set Ω(V _i ) where j∈[1,9], take SAD and SATD as block matching metrics, respectively, and calculate its rate-distortion performance through Lagrangian cost function and 5) according to the calculation result of step 4), extract the preset steganalysis feature to this frame group; 6) Repeat steps 2) to 5), successively carry out steganalysis feature extraction for all frame groups of the video to be tested; 7) Using a classifier based on steganalysis features, steganographic classification judgment is performed on each frame group in the video to be tested. 2. The method of claim 1, wherein step 4) rate-distortion performance and The calculation formula is: in, represents the reconstructed block with _Vi , represents the motion vector points to correspond to The prediction reference block of , λ is the Lagrange multiplier, represents the encoded motion vector required number of bits. 3. The method of claim 2, wherein step 5) extracts preset 4 types of sub-features to the frame group F _g , and finally merges to obtain 36-dimensional steganalysis features. 4. The method of claim 3, wherein step 5) comprises: a) Preprocessing: rate-distortion performance according to step 4) and Determine the minimum value of the elements in the set {J _SAD (m)|m∈Ω(V _i )}, denoted as Similarly, get b) Type 1 sub-feature extraction: each dimension of the type 1 sub-feature corresponds to the given k and Equal probability, defined as: Among them, the function The same below; c) Type 2 sub-feature extraction: Type 2 sub-features and J _SAD (V _i ) and The relative error between is related and is defined as: d) Type 3 sub-feature extraction: each dimension of the type 3 sub-feature corresponds to the given k and Equal probability, defined as: e) Type 4 sub-feature extraction: Type 4 sub-feature with J _SATD (V _i ) and The relative error between is related and is defined as: f) Final feature merging: The final 36-dimensional steganalysis feature is obtained by merging the above 4 types of sub-features and is defined as: