CN105721875B - A kind of video motion vector Stego-detection method based on entropy - Google Patents

Abstract

The invention discloses an entropy-based video motion vector steganographic detection method; the method uses a sliding window to select several motion vectors in a local area, and calculates the selected motion vectors in the horizontal component H, vertical component V, direction D, and length L The entropy of each of the four variables is the lowest 4 bits, and the 16-dimensional entropy value in a sliding window is obtained; by moving the sliding window in the video frame, multiple sets of entropy values are obtained, and the average value of the 16-dimensional entropy is calculated as a steganalysis feature. Through this method, the "entropy increase" anomaly caused by video motion vector steganography can be captured, and steganographic detection of motion vectors can be realized.

Description

A Steganographic Detection Method Based on Entropy in Video Motion Vector

technical field

The invention relates to the technical field of computer information security, in particular to an entropy-based video motion vector steganographic detection method.

Background technique

With the development of digital image information hiding and detection technology, information hiding technology based on digital video has also developed rapidly in recent years. In video steganography, steganography algorithms based on intra prediction, frequency domain and motion vector have appeared one after another. Among many steganography fields, the steganography algorithm based on motion vector has the characteristics of large hiding capacity and small image distortion. Therefore, motion vector steganography and the detection of steganography have become a research hotspot in the field of video information hiding.

In terms of concealment methods, the concealment algorithm uses a variety of strategies to reduce the image distortion caused by steganography: select the component with a larger absolute value in the motion vector for steganography; select the macroblock with a large difference between the reconstruction block and the reference block for motion vector concealment; Write; use wet paper coding to select low-cost channels for information embedding; use STC and wet paper coding to combine with each other to minimize the cost of embedding. The concealment of the algorithm is getting better and better, but the algorithm encoding generally uses the motion vector used in a single frame as a coding unit, which has not yet reached the minimum cost of the whole video macroblock, and does not consider the problem of maintaining the correlation of motion vectors. However, due to the use of one or more codes, the embedding algorithm is more complicated and not easy to implement.

In the face of various steganographic algorithms, the steganalysis detection of motion vectors has also begun to develop: use the correlation of adjacent differences to construct steganalysis features for steganalysis detection; use the correlation of multiple frames in space and time to detect motion vectors Steganography: use the motion vector to point to the optimal reference block characteristics for steganalysis, when the motion vector is adjusted, the optimal probability of the macro block decreases.

Current motion vector steganography detection algorithms are mainly based on motion vector correlation and local optimality, and there are two deficiencies: (1) Most steganographic detection methods using neighbor correlation construct detection features based on various neighbor probabilities, which have not been seen yet. to the feature construction method based on information entropy; (2) The steganographic detection method based on the local optimum of the motion vector needs to use the residual macroblock to reconstruct the optimal probability statistics. When the macroblock residual does not exist or the steganographic algorithm The detection ability of this kind of method is weak when the macroblock with small residual error is selected for steganography.

Contents of the invention

The object of the present invention is to provide an entropy-based steganographic detection method for video motion vectors. This method uses a sliding window to select a number of motion vectors in the local area, and calculates the entropy of the selected motion vectors in the horizontal component H, vertical component V, direction D, and length L. entropy value. By moving the sliding window in the video frame, multiple sets of entropy values are obtained, and the 16-dimensional entropy average value is calculated as the feature of steganalysis. Through this method, the "entropy increase" anomaly caused by video motion vector steganography can be captured, and steganographic detection of motion vectors can be realized.

A method for steganographic detection of video motion vectors based on entropy, comprising the steps of:

Step 1, the detector prepares cover non-steganographic samples, and performs steganographic embedding on the cover samples to obtain corresponding stego steganographic samples;

Step 2, extract the HVDL feature of each frame of the video, and add labels to the cover sample and stego sample to obtain the equivalent cover and stego training sample features; the HVDL feature is a 16-dimensional calculation based on motion vectors (H, V) Features; where H represents the horizontal component, V represents the vertical component, D is the direction of the motion vector, and L is the length of the motion vector;

Described extraction method comprises the following steps:

Step 2.1, decode a video frame containing the motion vector, and obtain the motion vector value of each macroblock in the frame; at this time, the video frame is converted into a combination of macroblocks of various sizes and their macroblock motion vectors (H, V) ;

Step 2.2, the motion vector video frame that step 2.1 obtains, selects sliding window; The motion vector in the sliding window is more than one, is regarded as a two-dimensional variable (H, V); The motion vector (H, V) this 2 Dimensional variables are expanded into 16-dimensional variables, and the entropy values of the 16-dimensional variables in the window are calculated, and finally multiple groups of entropy values are calculated by moving the sliding window;

The step 2.2 includes the following steps:

Step 2.2.1, set a sliding window with a size of m×n, where m and n are both positive integers; among them, m is the height of the sliding window, which represents m times the height of the maximum macroblock; n is the width of the sliding window , which represents that the width is n times the maximum macroblock width;

Step 2.2.2, for the motion vector (H, V) of the macroblock in the sliding window, convert the 2 variables into 16 variables by calculating the lowest 4 bits of the H component, V component, direction D and length L of the motion vector:

Among them, Q is a quantization function, taking the H component as an example, t represents the tth bit from the reciprocal of H, and the quantization methods of V, D, and L are the same; α is the angle formed by the motion vector and the horizontal right direction, α∈[0,2π), so D is the corresponding value from 0 to 15 after α is divided into 16 equal parts; Represents the rounded length of the motion vector;

Step 2.2.3, for each motion vector in the sliding window, it is regarded as the observation value of the motion vector (H, V) variable in the sliding window; based on step 2.2.2, the observation value of the 16-dimensional variable is obtained, Each variable takes the value 0 or 1; the entropy E of these 16 variables is:

E＝-(P ₀ log ₂ P ₀ +P ₁ log ₂ P ₁ )

Among them, P ₀ is the probability that the variable is 0, and P ₁ is the probability that the variable is 1;

In step 2.2.4, each sliding window can be calculated to obtain a set of 16-dimensional entropy values; the number of sliding windows is accumulated, and the entropy values of each dimension are accumulated;

Step 2.2.5, if the video frame has been scanned, then skip to step 2.3; otherwise, move the sliding window horizontally or vertically by one maximum macroblock unit, and a new window that is not exactly the same as the previously scanned window will be obtained. Sliding window, perform step 2.2.3;

In step 2.3, according to the cumulative number of sliding windows in step 2.2.5 and the cumulative sum of the entropy of each dimension, calculate the mean value of multiple sets of entropy value data to obtain the respective entropy mean values of 16 dimensions, and use it as a feature output; if Also have follow-up video frame, then return to step 2.1, if there is no follow-up video frame, then complete feature extraction;

Step 3, use SVM or other classifiers to perform feature training on the extracted features to obtain a training model;

Step 4, for any given video to be tested, also use the method of step 2 to extract HVDL features to construct detection features, use the training model trained in step 3 to predict whether the features of each frame are steganographic, and give Get the test result.

Further, in the step 2.2.1, the m=n=3.

Further, the maximum macroblock height and maximum macroblock width are 16 pixels.

The beneficial effect of the present invention is: a kind of entropy-based video motion vector steganography detection method; Adopt sliding window to select several motion vectors in the local area, calculate the selected motion vector in horizontal component H, vertical component V, direction D, and length The entropy of the lowest 4 bits of each of the four variables of L is obtained to obtain the 16-dimensional entropy value in a sliding window. By moving the sliding window in the video frame, multiple sets of entropy values are obtained, and the 16-dimensional entropy average value is calculated as the feature of steganalysis. Through this method, the "entropy increase" anomaly caused by video motion vector steganography can be captured, and steganographic detection of motion vectors can be realized.

Description of drawings

Fig. 1 is the motion vector steganography detection process of the present invention;

Fig. 2 is the HVDL feature extraction process of the present invention.

Detailed ways

As shown in Fig. 1, a kind of entropy-based video motion vector steganographic detection method comprises the following steps:

Step 1, the detector prepares cover non-steganographic samples, and performs steganographic embedding on the cover samples to obtain corresponding stego steganographic samples;

Step 2, extract the HVDL features of each frame of the video, and add labels to the cover samples and stego samples to obtain equivalent cover and stego training sample features;

As shown in Figure 2, the extraction method described comprises the following steps:

Step 2.1, decode a video frame containing the motion vector, and obtain the motion vector value of each macroblock in the frame; at this time, the video frame is converted into a combination of macroblocks of various sizes and their macroblock motion vectors (H, V) ;

Step 2.2, the motion vector video frame that step 2.1 obtains, selects sliding window; The motion vector in the sliding window is more than one, is regarded as a two-dimensional variable (H, V); The motion vector (H, V) this 2 Dimensional variables are expanded into 16-dimensional variables, and the entropy values of the 16-dimensional variables in the window are calculated, and finally multiple groups of entropy values are calculated by moving the sliding window;

The step 2.2 includes the following steps:

Step 2.2.1, set a sliding window with a size of m×n, where m and n are both positive integers; among them, m is the height of the sliding window, which represents m times the height of the maximum macroblock; n is the width of the sliding window , which represents that the width is n times the maximum macroblock width;

In said step 2.2.1, said m=n=3,

The maximum macroblock height and maximum macroblock width are 16 pixels.

Step 2.2.2, for the motion vector (H, V) of the macroblock in the sliding window, convert the 2 variables into 16 variables by calculating the H component, V component, direction D and length L of the lowest 4 bits of the motion vector:

Among them, Q is a quantization function, taking the H component as an example, t represents the tth bit from the reciprocal of H, and the quantization methods of V, D, and L are the same; α is the angle formed by the motion vector and the horizontal right direction, α∈[0,2π), so D is the corresponding value from 0 to 15 after α is divided into 16 equal parts; Represents the rounded length of the motion vector;

Step 2.2.3, for each motion vector in the sliding window, it is regarded as the observation value of the motion vector (H, V) variable in the sliding window; based on step 2.2.2, the observation value of the 16-dimensional variable is obtained, Each variable takes the value 0 or 1; the entropy E of these 16 variables is:

E＝-(P ₀ log ₂ P ₀ +P ₁ log ₂ P ₁ )

Among them, P ₀ is the probability that the variable is 0, and P ₁ is the probability that the variable is 1;

In step 2.2.4, each sliding window can be calculated to obtain a set of 16-dimensional entropy values; the number of sliding windows is accumulated, and the entropy values of each dimension are accumulated;

Step 2.2.5, if the video frame has been scanned, skip to step 3; otherwise, move the sliding window horizontally or vertically by one maximum macroblock unit, and a new window that is not exactly the same as the previously scanned window will be obtained Sliding window, perform step 2.2.3.

Step 2.3, according to the cumulative sliding window number in step 2.2.5, and the cumulative sum of the entropy of each dimension, calculate the mean value for multiple sets of entropy value data, obtain the 16-dimensional entropy mean value, and use it as a feature output; if there are Subsequent video frames, then return to step 2.1, if there is no subsequent video frame, feature extraction is completed.

Step 3, use SVM or other classifiers to perform feature training on the extracted features to obtain a training model;

Step 4, for any given video to be tested, also use the method of step 2 to extract HVDL features to construct detection features, use the training model trained in step 3 to predict whether the features of each frame are steganographic, and give Get the test result.

HVDL features: 16-dimensional features calculated based on motion vectors (H, V), where H represents the horizontal component, V represents the vertical component, D is the direction of the motion vector, and L is the length of the motion vector.

Claims (3)

1. a kind of video motion vector Stego-detection method based on entropy, it is characterised in that：Include the following steps：

Step 1, tester gets out the non-steganography samples of cover, and steganography insertion is carried out on cover samples, obtains corresponding Stego steganography samples；

Step 2, label is added per frame HVDL features, and to cover samples and stego samples in extraction video, obtains equity Cover and stego training sample features；The HVDL features are 16 dimensional features being calculated according to motion vector (H, V)； Wherein H indicates that horizontal component, V indicate vertical component, and D is the direction of motion vector, and L is the length of motion vector；

The extracting method includes the following steps：

Step 2.1, a video frame for including motion vector is decoded, the motion vector numerical value of each macro block in the frame is obtained；This When video frame be converted into each size dimension macro block and its macroblock motion vector (H, V) combination；

Step 2.2, to step 2.1 obtain include motion vector video frame, choose sliding window；Movement arrow in sliding window More than one is measured, regards a two-dimentional variable (H, V) as；By motion vector (H, V), this 2 dimension variable is extended to 16 dimension variables, And the 16 dimension variable entropy in calculation window, it finally moves sliding window and calculates multigroup entropy；

The step 2.2 includes the following steps：

Step 2.2.1 sets a size as the sliding window of m × n, and wherein m and n are positive integer；Wherein, m is sliding window Open height represents m times of maximum macro block height；N is the width of sliding window, represents the width as maximum macro block width N times；

Step 2.2.2, for the motion vector (H, V) of macro block in sliding window, by calculating the H components of motion vector, V divides 2 variables are converted to 16 variables by the respective minimum 4bit of amount, direction D and length L：

Wherein, Q is quantization function, by taking H components as an example,T represent take H inverse t-th of bit, V, D, the quantization method of L is identical；The angle that α is formed for motion vector with horizontal right direction, α ∈ [0,2 π), therefore D It is that α is divided into corresponding 0 to 15 numerical value after 16 deciles；Represent the rounding length of motion vector；

Step 2.2.3, for each motion vector in sliding window, all regard as motion vector in this sliding window (H, V) the observation of variable；Based on step 2.2.2, the observation of 16 dimension variables, each variable-value 0 or 1 have been obtained；This 16 changes The entropy E of amount is：

E=- (P₀log₂P₀+P₁log₂P₁)

Wherein, P₀The probability that i.e. variable is 0, P₁I.e. variable is 1 probability；

One group of 16 dimension entropy can be calculated in step 2.2.4, each sliding window；Cumulative calculation sliding window quantity, and tire out Meter is per one-dimensional entropy；

Step 2.2.5, if having scanned through this video frame, jump procedure 2.3；Otherwise, by one sub-level of sliding window or hang down The maximum macro block unit of dynamic one of translation, can obtain one be scanned the new sliding window of window not exclusively repeatedly before this, Execute step 2.2.3；

Step 2.3, according to the accumulative sliding window quantity in step 2.2.5, and its sum of the accumulation of each dimension entropy, to multigroup entropy Data are averaged, and obtain the respective entropy mean value of 16 dimensions, and export it as feature；If also subsequent video frame, step is returned to Rapid 2.1, if without subsequent video frame, complete feature extraction；

Step 3, using SVM or other graders, the training of feature is carried out to the feature of extraction, obtains training pattern；

Step 4, same to be detected using the method extraction HVDL latent structures of step 2 for any one given video to be measured The training pattern that training obtains in step 3 is made whether obtained every frame feature the prediction of steganography, provides detection by feature As a result.

2. a kind of video motion vector Stego-detection method based on entropy according to claim 1, it is characterised in that：It is described Step 2.2.1 in, the m=n=3.

3. a kind of video motion vector Stego-detection method based on entropy according to claim 1, it is characterised in that：It is described Maximum macro block height and maximum macro block width be 16 pixels.