CN103888773A - Video steganography analysis method based on mutual information and motion vectors - Google Patents

Abstract

The invention provides a video steganography analysis method based on mutual information and motion vectors, comprising steps of adopting an additive Gaussian noise superposition steganography method in a process of establishing a model, using entropy of the motion vector in a prediction frame in a compressed video flow to measure the influence on the spatial correlation, utilizing a difference histogram of a horizontal direction and a vertical direction of the motion vector between adjacent prediction frames and mutual information to measure the influence on the space correlation, and determining whether the steganography exists according to the classification test of a support vector machine. In the light of the steganography of the motion vector, the method provided by the invention can achieve better detection effect.

Description

A kind of based on mutual information and the hidden analytical method of writing of motion vector video

Technical field

The present invention relates to the hidden analytical method of writing of a kind of video, relate in particular to a kind of based on mutual information and the hidden analytical method of writing of motion vector video.

Background technology

Increasingly mature along with the constantly universal and digital multimedia application technology of the Internet, particularly network flow-medium business development, makes audio frequency, video, the multimedia transmission such as image and exchange are very efficient and convenient.The work that shares to people of information resources, life, amusement have brought unlimited facility, particularly, along with the continuous of Video Applications popularized, various video files on network are constantly occurred; Meanwhile, Information Hiding Techniques is also at development, and the experience of image information concealing technology accumulation is also for application video provides condition for carrier carries out Information hiding.

Research video Steganalysis tool is of great significance.First, video Steganalysis can be contained the abuse of video steganography, along with the progress of video Steganalysis, detection algorithm accuracy rate, detection speed for various video steganographies are all significantly improving, and this has just improved the hidden supervision of writing of video to a great extent.Secondly, research video Steganalysis is also the inevitable requirement of capturing information security research field commanding elevation, video Steganalysis is field newer in Information hiding, but be also the important directions of future development simultaneously, only have and grasped advanced video Steganalysis, could guarantee to obtain initiative safely with in military affairs, intelligence activity at national information.Finally, research video Steganalysis can reach the facilitation to video steganography, finds the deficiency of existing video steganographic algorithm by video steganalysis, then for existing shortcoming, improves steganographic algorithm, just can reach that better video is hidden to be write.

Motion vector is hidden to be write compared with traditional spatial domain or frequency domain steganography method, and the hidden disguise of writing of motion vector is better.Compared with image latent writing analytical technology, the development of video steganography is slower, and Most scholars becomes video modeling the combination of a series of rest images, and the embedding of information can be modeled as to signal and add that average is zero Gaussian noise.This model can be applied to that motion vector is hidden to be write equally.Motion vector is carried out to the hidden time-space domain statistical nature that can affect frame of video of writing.Therefore can carry out steganalysis by the more hidden variation of writing the vectorial statistical nature that seesaws.

Summary of the invention

Technical problem: the object of this invention is to provide a kind of for the hidden analytical method of writing of motion vector.Motion vector in video code flow is that the motion estimation algorithm by Video coding obtains, and it is related to time continuity and the flatness of video pictures, is that video compression is laid special stress on protecting object.Through coding or again after encoding and decoding, motion information data changes very little.Due to motion vector reflection be the moving displacement information of optimum Match macro block in predicted macro block and reference frame in current encoded frame, have nothing to do with the particular content of macro block.If information is embedded in motion vector, there is very strong disguise.The method that the application of the invention proposes can effectively improve the verification and measurement ratio of this type of steganography method.

Technical scheme: write the additive process that process model building is additive Gaussian noise to hidden, utilize the entropy of predictive frame intraframe motion vector in compressing video frequency flow to measure the hidden impact of writing spatial coherence, utilize histogram of difference and the mutual information of adjacent predictive frame interframe movement vector horizontal and vertical two directions to measure the hidden impact of writing temporal correlation simultaneously.Finally judge whether to exist hidden writing by the classification and Detection of SVMs.

The present invention comprises following steps:

1, the common signal channel between transmit leg and recipient extracts the motion vector in compressing video frequency flow, obtains binary data stream according to the correlation judgement of motion vector in intraframe motion vector and its neighborhood, calculates the entropy of this binary data stream.By video stream data is carried out to partial decoding of h, can obtain motion vector and the relevant information thereof of encoder motion prediction frame.When object video code stream successively decomposes macroblock layer, by judging whether tell present frame is P inter-coded macroblocks and the B interframe alternating binary coding macro block based on encoder motion prediction.In P frame, can calculate forward motion vector, and in B frame, can calculate two motion vectors of forward and backward.Draw corresponding data by these computings.

2, calculate predictive frame interframe mutual information I.

3, calculate predictive frame frame-to-frame differences histogram front Fourth-order moment, i.e. average μ, variances sigma ², degree of bias ζ, kurtosis κ.

4, the characteristic of division extracting from training sample database, as the input of selected grader, uses the penalty factor of grid search method Automatic-searching optimum and the parameter σ of kernel function in cross-validation process.

5, the input using the characteristic vector of video sample to be measured as the grader having trained, classifies, and obtains classification results.

Adopt interframe mutual information and binary data stream entropy to measure the hidden impact of writing carrier video temporal correlation.Adopt mutual information and the histogrammic average of predictive frame frame-to-frame differences, variance, the degree of bias, kurtosis to represent interframe feature.The motion vector correlation that present frame is adjacent frame changes after writing motion vector is hidden, and from the original approximate identical difference that becomes, the motion vector degree of restraint that has reacted adjacent two frames diminishes.And this degree of restraint can represent with the mutual information of adjacent two frames.The entropy of the binary data stream of feature by calculating arbitrary predictive frame obtains in frame.

Accompanying drawing explanation

Fig. 1 is system block diagram of the present invention.

Fig. 2 is flow chart of the present invention.

Fig. 3 is motion vector analysis process figure of the present invention.

Fig. 4 is steganalysis ROC curve chart of the present invention.

Embodiment

Below in conjunction with accompanying drawing and embodiment, the present invention is described in further detail.

As shown in Figure 1, be system block diagram of the present invention.Video steganalysis is exactly the technology that the Information hiding take video as carrier is analyzed.Hidden write process be exactly transmit leg under the effect of key K, utilize embed algorithm secret information is write to video, pass to recipient by common signal channel, recipient, under the effect of key K, utilizes extraction algorithm that secret information is extracted.Video steganalysis is exactly on common signal channel, the video of transmission to be detected, and analyzes and judges in video, whether there is Information hiding.

The present invention is directed to existing motion vector steganalysis algorithm and make improvement.First, adopt mutual information and the histogrammic average of predictive frame frame-to-frame differences, variance, the degree of bias, kurtosis to represent interframe feature.The motion vector correlation that present frame is adjacent frame changes after writing motion vector is hidden, and from the original approximate identical difference that becomes, the motion vector degree of restraint that has reacted adjacent two frames diminishes.And this degree of restraint can represent with the mutual information of adjacent two frames.If two frame motion vectors are separate, mutual information is 0; If two frame motion vector correlations are stronger, mutual information value is larger.Therefore can extract consecutive frame mutual information feature and weigh the interframe characteristic of motion vector.Not do not embed while hiding Info, in frame-to-frame differences, except the edge of moving region has compared with complex texture, other most of region is level and smooth, illustrates and between consecutive frame, has strong correlation.Have in a frame and embedded information and need only in adjacent two frames, in frame-to-frame differences, not only the edge of moving region has compared with complex texture, and the Texture complication in other region also obviously strengthens, and illustrates that the related receptor of consecutive frame has arrived weakening.

The macro block of supposing one 16 × 16 is a motion vector.First judge the correlation of current motion vector and its neighborhood motion vector, if relevant, be designated as 1, otherwise be designated as 0.Here neighborhood is defined as 8 motion vectors that surround current motion vector.Arbitrary predictive frame is calculated to binary data stream.In the time not embedding information, the correlation between motion vector is stronger, must obtain more 1.And after motion vector is modified, destroyed motion vector and the correlation of motion vector around, cause 1 number to reduce.Entropy can be used for weighing the randomness of a given distribution, and by binary data stream, by every 8 one group, the entropy of definition motion vector binary data stream is:

$\mathrm{En}=-\underset{i=1}{\overset{356}{\mathrm{\Σ}}}\left(p\left(x_i\right)\mathrm{lgp}\left(x_i\right)\right)$

The entropy of revising the binary data stream after motion vector is less than the entropy of original binary data stream.

Suppose to have two any predicted video frame k1 and k2 (k1 ≠ k2), the entropy of k1, k2 is:

$H\left(k1\right)=-\underset{i}{\mathrm{\Σ}}{p}_{k1}\left(i\right)\log p_{k1}\left(i\right)$

$H\left(k2\right)=-\underset{j}{\mathrm{\Σ}}{p}_{k2}\left(j\right)\log p_{k2}\left(j\right)$

The combination entropy of two predicted video frame:

$H\left(X_{k1}{X}_{k2}\right)=-\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{p}_{k1}{p}_{k2}(i,j)\log p(i,j)$

Wherein P _k1, P _k2, P _k1k2be respectively probability distribution and the joint probability distribution of k1, k2.

Between definition frame, motion vector mutual information is as follows:

I(X _k1;X _k2)＝H(X _k1)-H(X _k1/X _k2)

The mutual information that obtains frame k1 and k2 is:

I(X _k1;X _k2)＝H(X _k1)+H(X _k2)-H(X _k1X _k2)

Frame-to-frame differences histogram is discrete function, is designated as h (d _k)=n _k/ n, wherein d _k∈ [M+1, M-1] is k difference, and M is the maximum of motion vector, k ∈ [1,2M-1]; n _kin difference image, to equal d _kthe number of difference, n is the sum of difference in difference image.Visible, h (d _k) provide difference d _kthe probability Estimation value occurring.Here, according to the front Fourth-order moment of the poor histogram calculation difference image of predictive frame motion vector, i.e. average μ, variances sigma ², degree of bias ζ, kurtosis κ, each computational methods are as follows:

$u=\underset{i=1}{\overset{2m-1}{\mathrm{\Σ}}}{d}_{i}h\left(d_i\right)$

${\mathrm{\σ}}^2=\underset{i=1}{\overset{2m-1}{\mathrm{\Σ}}}{(d_i-u)}^{2}h\left(d_i\right)$

$\mathrm{\ζ}=\frac{\underset{i=1}{\overset{2m-1}{\mathrm{\Σ}}}{(d_i-u)}^{3}h\left(d_i\right)}{{\mathrm{\σ}}^3}$

$\mathrm{\κ}=\frac{\underset{i=1}{\overset{2m-1}{\mathrm{\Σ}}}{(d_i-u)}^{4}h\left(d_i\right)}{{\mathrm{\σ}}^4}$

Because SVMs (Support Vector Machine, SVM) can be classified to characteristic vector preferably in high dimensional nonlinear space, less to sample size and quality dependence, therefore adopt SVM as grader.In SVM method, use different inner product functions, just can realize the multiple learning algorithms such as multinomial approaches, RBF (Radial Basic Function, RBF).Use the SVMs of RBF core to be widely used in pattern recognition.The Model Selection of this type of SVM depends on two parameters, first penalty factor, and it two is nuclear parameter γ.This method adopts grid search method to determine penalty factor and nuclear parameter γ.Penalty factor and nuclear parameter γ get respectively M and N value, to the combination of M × N (C, γ), train respectively different SVM, estimate again its generalized recognition rate, thereby in the combination of M × N (C, γ), be promoted a combination that discrimination is the highest as optimized parameter.Adopt RBF as kernel function:

$K(x_i,x_j)=\exp (-\frac{{\left|\right|x_i-x_j\left|\right|}^2}{2{\mathrm{\γ}}^2})$

Motion vector has horizontal component and vertical component, calculates respectively both front Fourth-order moment, can obtain like this 8 characteristic values.The entropy that adds predictive frame mutual information and binary data stream obtains 10 characteristic values, composition characteristic vector altogether.After characteristic vector is constructed, whether in video sequence, containing hides Info just becomes 2 classification problems.Algorithm steps as shown in Figure 2.

Fig. 3 is motion vector analysis process figure of the present invention.Video stream data can obtain motion vector information by some read means.As long as find forward predicted frame, just can obtain forward motion vector, obtain bi-directional predicted frames and can obtain the motion vector on forward and backward both direction.Carry out certain calculating by the motion vector that these are obtained and just can obtain our required information.

Beneficial effect of the present invention can further be verified by emulation:

Download 50 sections of CIF(352 × 288 from the Internet) video flowing of form, YUV adopts 4:2:0.These 50 sections of video flowings are divided into 240 sections by every section of 30 frames, and first 120 sections are used for training SVM, and latter 120 sections are used for testing.These 120 sections of test video levellings are divided into 4 groups, adopt three kinds of different motion vector steganographic algorithms to carry out hidden writing, the 4th group is not carried out hidden writing to it.Receiver operating characteristics curves is a curve between sensitivity and feature.The height of verification and measurement ratio depends primarily on the method for classification, and result can form accepts operating characteristic curve (ROC), and abscissa is false alarm rate, and ordinate is verification and measurement ratio.

Steganalysis ROC curve chart of the present invention from Fig. 4: be 0 o'clock at false alarm rate, verification and measurement ratio of the present invention is more than 60%.To obviously be better than the other two kinds of algorithms of document to detection the present invention of compressed video.These two kinds of algorithms, not for compressed video, can affect the detection performance of its algorithm after video compression.Owing to the present invention is directed to predictive frame motion vector, algorithm complex is lower, can find out that from ROC curve of the present invention the present invention still has higher verification and measurement ratio.

Claims (3)

1. based on mutual information and the hidden analytical method of writing of motion vector video, it is characterized in that comprising following steps:

Step 1: the common signal channel between transmit leg and recipient extracts the motion vector in compressing video frequency flow, obtain predictive frame binary data stream according to the correlation judgement of motion vector in motion vector in arbitrary predictive frame and its neighborhood, calculate the entropy of this binary data stream;

Step 2: calculate predictive frame interframe mutual information;

Step 3: calculate the front Fourth-order moment of predictive frame frame-to-frame differences histogram;

Step 4: the characteristic of division extracting from training sample database, as the input of selected grader, uses the penalty factor of grid search method Automatic-searching optimum and the parameter σ of kernel function in cross-validation process;

Step 5: the input using the characteristic vector of video sample to be measured as the grader having trained, classify, obtain classification results.

2. according to claim 1 a kind of based on mutual information and the hidden analytical method of writing of motion vector video, it is characterized in that:

In described step 1, the entropy that calculates this binary data stream is:

$\mathrm{En}=-\underset{i=1}{\overset{256}{\mathrm{\Σ}}}\left(p\left(x_i\right)\mathrm{lgp}\left(x_i\right)\right);$

In described step 2, calculating predictive frame interframe mutual information is:

I(X _k1;Xk2)＝H(X _k1)+H(X _k2)-H(X _k1X _k2)；

In described step 3, before calculating predictive frame frame-to-frame differences histogram, Fourth-order moment is: average μ, variances sigma ², degree of bias ζ, kurtosis κ.

3. according to claim 2 a kind of based on mutual information and the hidden analytical method of writing of motion vector video, it is characterized in that, the characteristic vector of described video sample to be measured is: the horizontal component of calculation of motion vectors and the front Fourth-order moment of vertical component respectively, obtains 8 characteristic values; The entropy that calculates predictive frame mutual information and binary data stream, obtains 2 characteristic values, and totally 10 characteristic values form the characteristic vector of video sample to be measured.

Images