CN105872555A - Steganalysis algorithm specific to H.264 video motion vector information embedment - Google Patents
Steganalysis algorithm specific to H.264 video motion vector information embedment Download PDFInfo
- Publication number
- CN105872555A CN105872555A CN201610177535.5A CN201610177535A CN105872555A CN 105872555 A CN105872555 A CN 105872555A CN 201610177535 A CN201610177535 A CN 201610177535A CN 105872555 A CN105872555 A CN 105872555A
- Authority
- CN
- China
- Prior art keywords
- dtri
- motion vector
- video
- order difference
- rsqb
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/46—Embedding additional information in the video signal during the compression process
- H04N19/467—Embedding additional information in the video signal during the compression process characterised by the embedded information being invisible, e.g. watermarking
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N17/00—Diagnosis, testing or measuring for television systems or their details
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Signal Processing (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
Abstract
The invention discloses a steganalysis algorithm specific to H.264 video motion vector information embedment. The algorithm aims at rapidly and effectively carrying out steganalysis to an H.264 video, and especially has relatively high detection rate specific to an embedding algorithm which keeps statistic characteristics unchanged and a low embedding rate relatively well. The technical scheme used by the invention is that motion vectors are accurately extracted; high order differential characteristics are designed on the basis of first order differential characteristics; according to this kind of characteristics, steganalysis can be carried out relatively well on a steganography algorithm which compensates vector statistic characteristics; multiple simple classifiers are trained by using integrated learning thought; an integrated steganography classifier based on FLD is designed; and the integrated classifier is relatively greatly improved in classification speed; and compared with single same classifiers, the integrated classifier has higher classification accuracy.
Description
Technical field
The invention belongs to Information Hiding Techniques field, be specifically related to a kind of for the embedding of H.264 video motion vector information
Steganalysis algorithm.
Background technology
Steganography and steganalysis are the main contents of Information Hiding Techniques, important as safe information transmission of steganography
Means, can apply to the aspects such as military affairs, information, national security, provide technical guarantee for communication security.Owing to video carrier has
Having larger data redundancy, human eye is to the feature such as moving image sensitivity is relatively weak so that digital video is that Large Copacity is hidden
Communication provides essential condition.Research about video concealing technology at present constantly heats up, and network utilizes digital video conduct
Carrier carries out the software of Information hiding to be occurred the most in succession.Video information steganalysis must be weighed as an important problem
Depending on and pay close attention to.
Motion prediction and compensation technique utilize frame of video dependency over time and space, dramatically saves on video counts
According to memory space, be widely used in all kinds of video compression standard.Motion vector is as in compression Video coding
Important component part, there is higher stability, image quality can't be caused big impact by a small amount of change, the suitableeest
Cooperation is the position of Information hiding, and quite a few video steganographic algorithm is exactly to select the motion meeting certain condition to vow at present
Amount, information is embedded in the amplitude of motion vector, phase angle, the direction of motion, horizontal component and vertical component or predictive value vector it
In, this kind of information steganography algorithm based on motion vector has higher safety, not sentience and information and embeds and extract
Real-time, with the feature such as video encoding standard combines.
H.264 as a kind of new, widely used video compression standard, there is compared with conventional standard support less
Piecemeal, predict more accurately, compressed encoding can comprise more motion vector, H.264 this make, and normal video is very
It is suitable for motion vector as embedded location, hides information.Such Information Hiding Techniques is due to light to the modulation of motion vector
Micro-, therefore the least on the impact of frame of video quality, for frame of video itself, there is not specific regularity, therefore cannot be from rebuilding figure
As reflecting the existence of hiding information on frame, traditional statistical analysis algorithms based on video frame pixel cannot detect point effectively
Separate out this type of video steganographic algorithm.
Currently, Chinese scholars has been carried out substantial amounts of digital video steganography and the rational approach of steganalysis,
The most actively research is practical, safety is high, the Steganalysis of perfect in shape and function in a lot of countries.But it is the most relevant
Video Steganalysis, particularly to combining a new generation's video compression standard, for the steganalysis algorithm of motion vector
Research or ratio are relatively limited.
At present, the Steganalysis for image develops very fast.Since the mid-90 in 20th century, propose both at home and abroad
Image latent writing analytical technology has tens kinds more than, has nothing in common with each other in direction, and wherein the overwhelming majority belongs to statistics characteristic analysis method.
As: Fridrich etc. propose a kind of based on single order with the steganalysis algorithm of second order distributed feature for jpeg image steganography,
And by it as a comparison jpeg image steganographic algorithm embed mechanism appraisal tool;Jackson etc. extract feature at Farid algorithm
On the basis of, devise Stego-detection system (CIS) based on artificial immunity;It is similar that Avcibas selects between bit-planes
Degree sets up characteristic vector, then sets up grader by multiple linear regression model, but this method is only applicable to LSB steganography
The detection of algorithm;Westfeld is according to the system to (pairs of value) of the value in carrier image before and after hiding information at LSB
Meter property difference devises chi-square criterion method.Domestic Wang Shuozhong, Zhang Xinpeng et al. are to regard to digital picture steganography and steganalysis
Technology is contrasted careful analysis, has published " Steganography system and steganalysis " book in 2005.In book, ratio is in greater detail
Describe some common steganography methods, steganalysis method and anti-steganography method.
Compared with image latent writing analytical technology, the development of video Steganalysis is later and slower.Existing known steganography is divided
Analysis algorithm mainly has Su Yu very to wait the steganalysis algorithm for MSU video hide tools of proposition, and this algorithm is from embedding data
After the blocking effect change that causes start with, for not having the video of embedding data, the distribution of the blocking effect of its 16x16 block and 8x8 is basic
Equally, when there being data to embed, the blocking effect of its 16x16 is apparently higher than 8x8 blocking effect, on this basis, resets a thresholding
Value, carries out Stego-detection.But this algorithm is the special Stego-detection for MSU, to the video of additive method steganography inapplicable.
The steganalysis method based on collusion attack that Udit Budhiad etc. [8] propose, utilizes collusion attack to remove the secret of target frame
Confidential information, implements collusion attack to given video, as long as video rate of change is very slow and implements the window L choosing of collusion attack
Take the approximation that appropriately just can obtain original video, thus estimate the secret information of embedding.This algorithm is owing to being to utilize phase
The time redundancy of adjacent interframe, when video change is violent, and when having more global motion, Detection accuracy is decreased obviously.
Julien.S.Jainsky etc. propose asymptotic memoryless detection method.Utilize present frame and the motion vector of frame, employing before and after it
Video motion interpolation reconstruction video, then compares with former video, analyses whether, finally by ARE arbiter, the information that carried out
Hide.This algorithm the most only make use of the time redundancy of video, in the case of video not exclusively embeds, and particularly current video frame
The information of hiding, and when before and after it, consecutive frame is not hidden, it is possible to present frame is accurately detected, but hidden for Large Copacity information
Hide the whole detection accuracy rate not the biggest raising of video sequence.Su Yu very waits propose to utilize the first-order difference of motion vector straight
The feature of side's figure, calculates its corresponding statistic, and the algorithm embedded for being directly based upon motion vector information has preferably inspection
Survey efficiency, but as motion vector is compensated so that it is rectangular histogram keeps original characteristic, then cannot be carried out effectively detecting.Yu
The steganalysis algorithm embedded for motion vector that Deng proposes, this algorithm utilizes the corresponding statistic of second differnce as spy
Levy, carry out corresponding steganalysis.This steganographic algorithm keeping first-order characteristics has preferable analytical effect, but for single order
The steganographic algorithm of the equal retention performance of second order cannot obtain preferable verification and measurement ratio.
The detection of video steganalysis is at the early-stage, and related algorithm is less, and it is studied oneself is extremely urgent.Existing
Detection algorithm need to improve in the suitability and Detection accuracy, is therefore specifically designed for H.264 normal video motion vector embedding
The detection algorithm entered has certain practical value.
Summary of the invention
In order to solve the problems of the prior art, the present invention proposes a kind of for the embedding of H.264 video motion vector information
Steganalysis algorithm, it is possible to fast and effeciently H.264 video is carried out steganalysis, especially for can preferably keep system
The embedded mobile GIS of the meter constant and low embedding rate of characteristic has higher verification and measurement ratio.
In order to realize object above, the technical solution adopted in the present invention is: comprise the following steps:
1) motion vector extracts: is decoded video, obtains the motion vector of each macro block, as motion vector difference
The data of feature extraction;
2) to step 1) data that obtain carry out motion vector Differential Characteristics extraction:
2.1) first-order difference feature extraction:
First-order difference distortion factor E of calculating motion vector:
Wherein, kurtosis (Si) it is the kurtosis value of motion vector first-order difference, SiFor motion vector first-order difference,
▽Si=Si-Si+1, SiBeing the value of motion vector in i-th block, h [-2], h [-1], h [0], h [1], h [2] represent S respectivelyi
Probability mass function;
Calculate the average of single order vector difference
Wherein, q is the amplitude range of motion vector first-order difference, hS[n] is SiProbability mass function;
Calculate the variance of single order vector difference
Wherein, μsFor the average of motion vector, hS[n] is SiProbability mass function;
By E,As first-order difference feature;
2.2) second differnce feature and third order difference feature extraction:
The center mass function of calculating motion vector second differnce:
Wherein,Represent respectively2Xi, second differnce2S、▽2The probability mass function of η
Fourier's series, i is the number of block in a frame, and N is the number of block in frame of video;
The variance of calculating second differnce:
Wherein, μsFor the average of motion vector,It it is second differnce2The probability mass function of S, q is motion vector
The amplitude range of first-order difference;
Calculating second differnce and the kurtosis value of third order difference:
Wherein,WithRepresent motion vector second differnce and the average of third order difference respectively,Representing motion vector second differnce and the variance square of third order difference respectively, N is frame of video
The number of middle piece;
By second differnce statisticAnd third order difference statisticAs higher difference feature;
3) FLD integrated classifier is used according to first-order difference feature and higher difference feature, the video segment of input to be carried out
Detection judges, statistics is wherein judged as that the number of steganography and normal video frame, to judge whole video segment, i.e. completes steganography and divides
Analysis algorithm.
Described step 1) in use FFmpeg decoder video is decoded, decode_slice_header ()
Function obtains the frame number of current decoded frame, h1_motion () function obtains the type of each macro block, at mc_dir_
Part () function obtains the motion vector of each macro block, and exports in TXT text, extract as motion vector Differential Characteristics
Data.
Described step 2.1) in the computing formula of kurtosis value of motion vector first-order difference as follows:
Wherein, μ (Si) it is the average of motion vector first-order difference, σ4(▽Si) represent motion vector first-order difference side
Difference square, N is the number of block in frame of video.
Described step 2.1) in introduce first-order difference function center mass feature reflect motion vector first-order difference
Energy distribution situation, by calculating center mass feature C after first-order difference probability mass function Fourier transform1(H
[m]), reflect that vector first-order difference is in frequency domain energy variation situation:
Wherein, H [i] represents, T represents, i represents the number of block in a frame.
Described step 2.1) in the first-order difference of motion vector be defined as follows:For carrier video
The first-order difference Distribution value of motion vector meets the super-Gaussian distribution that peak value is 0, the horizontal component after embedding information and vertical point
The motion vector first-order difference of amountWithIt is expressed as:
Wherein,WithIt is respectively horizontal component and the embedding information of vertical component of first-order difference,WithPoint
Wei horizontal component and the first-order difference of vertical component.
Described step 2.2) defined in the second differnce of adjacent motion vectors be:
▽2Si=Si-▽Si+1
After information embedding, the horizontal and vertical component second differnce of motion vector is expressed as:
Wherein,WithIt is respectively horizontal component and the embedding information of vertical component of second differnce,WithIt is respectively horizontal component and the first-order difference of vertical component.
Described step 2.2) in third order difference be defined as follows:
Described step 3) in FLD integrated classifier include several base learners, each base learner Bl, l=
1 ..., L, is Rd→ { mapping of 0,1}, wherein 0 represents carrier frame, and 1 represents steganography frame, RdFor all full dimensional characteristics, often
The decision-making value of one base learner is adjusted in the case of waiting priori, and the error rate minimizing training is:
Wherein PFA,PMDIt is the probability of false-alarm and missing inspection respectively.
Each described base learner uses Fisher linear discriminant to be used as learning tool.
The lth base learner of described FLD integrated classifier is in training setUpper training
Obtain, whereinIt is the subset randomly choosed, simultaneouslyBe from set 1 ..., NtrnAdopt in }
The bootstrapping sample that sample obtains,The characteristic vector of each base learner is expressed as:
WhereinIt is the average of each class: Being the Scatter Matrix in class, wherein λ is steady
Determine parameter to ensure SW+ λ I is positive definite.
Compared with prior art, present invention extraction accurate to motion vector, by simplifying FFmpeg and revising, only
Retain h.264 decoded portion therein, and increase the output function of motion vector information, while decoding quickly and accurately
Realize the extraction of motion vector, highly versatile.Higher difference feature is devised, this category feature on the basis of first-order difference feature
The steganographic algorithm having vector statistics feature to compensate can preferably carry out steganalysis, and detection efficiency is high, versatility, stable
Property is strong.Video image characteristic quantity is big, by single complex classifier, at support vector machine or Method Using Relevance Vector Machine
The detection time needed for reason is the longest, and with simple classification device such as FLD, then Detection results is poor, in time efficiency and accuracy rate
Between be difficult to seek an equilibrium point, utilize the thought of integrated study, multiple simple classification devices be trained respectively, design one
Planting integrated steganography grader based on FLD, integrated classifier is not only greatly improved on classification speed, and more single than using
The same category utensil of one has higher classification accuracy.The steganalysis algorithm that resisted motion Vector Message of the present invention embeds,
And realize its software system, on the basis of feature extraction and classifier design, it is designed to resist what motion vector information embedded
Steganalysis algorithm, and with H.264 carrier for detection object, it is achieved steganalysis software system.
Accompanying drawing explanation
Fig. 1 is that motion vector of the present invention extracts flow chart;
Fig. 2 is integrated classifier of the present invention training and detection;
Fig. 3 is work model figure of the present invention;
Fig. 4 a is the first-order difference scattergram of the motion vector of normal video;Fig. 4 b is the one of the motion vector of concealed video
Jump divides scattergram.
Detailed description of the invention
Below in conjunction with specific embodiment and Figure of description, the present invention is further explained.
The video steganalysis of the present invention, based on H.264/AVC standard, is specifically designed for H.264 motion vector in video
Embedding grammar carries out steganalysis, and key technology is the most H.264 video to be carried out steganalysis, especially pin
The embedded mobile GIS that can preferably keep the constant and low embedding rate of statistical property is had higher verification and measurement ratio, sees Fig. 3, specifically
As follows:
1) motion vector extracts:
Carrying out steganalysis for motion vector, the accurately extraction of motion vector is the premise of steganalysis, for H.264
For code stream, motion vector to be extracted, first have to it is decoded, the encoder meeting H.264 coding standard at present has very
Many, different encoders, it is generally required to the decoder of its correspondence decodes, therefore selects a versatility good, fireballing
H.264 encoder is particularly significant, and FFmpeg is a cross-platform video increased income and audio codec, it provides at present
The decoding scheme of main flow Voice & Video, so that the code stream of different coding device is had preferable versatility, sees Fig. 1,
The present invention uses FFmpeg to be decoded video, and FFmpeg is powerful and highly versatile, but it is mainly in Linux platform
Lower exploitation, in order to make its stable operation under windows system, again it is compiled under windows environment, with
Time code is simplified, only remain decoded portion the most H.264, add on this basis motion vector output mould
Block, obtains the frame number of current decoded frame in decode_slice_header () function, obtains in h1_motion () function
The type of each macro block, obtains the motion vector of each piece in mc_dir_part () function, these information is exported
In TXT text, the initial data calculated as motion vector Differential Characteristics.
2) motion vector Differential Characteristics extracts:
Feature is the key point of steganalysis, and feature the most effectively direct relation the height of verification and measurement ratio, based on motion
The information of vector embeds can regard addition noise in the horizontal component and vertical component of motion vector as, and it respectively can be with table
Show as follows:
WithBeing the value of horizontal and vertical movement vector component in i-th block respectively, N is the number of block in a frame,
WithBeing the information embedded, its probability mass function PMF is expressed as follows,P is
The embedding rate of information in algorithm, k represents the modified values in algorithm to carrier vector, the shadow to carrier after embedding information as k=1
Ringing minimum, embedded mobile GIS based on motion vector generally, is the S selecting to meet certain conditioniEmbed information, such as Si
Amplitude, phase angle etc.,Represent the motion vector after embedding information.Steganographic algorithm based on motion vector is divided
Analysis, essence finds out the difference of the motion vector of carrier and concealed video exactly, and therefore we to design corresponding statistic and come anti-
S after answering information to embediDistortion situation, the present invention carries out Difference Calculation to motion vector, and with the system of these difference result
Metering is as detection feature;
(1) first-order difference feature:
There is between the motion vector that space is adjacent good dependency, use in same frame the one of neighboring block motion vector
Jump divides to embody this correlation properties.The first-order difference of motion vector is defined as follows:Due to dependency,
First-order difference Distribution value for carrier video motion vector meets the super-Gaussian distribution that peak value is 0.Shown in its scattergram 4a:
The motion vector first-order difference of the horizontal and vertical component after embedding information can be expressed as:
According to ηiProbability mass function, ηiPMF be, It is hereby achieved thatDistribution, such as Fig. 4 b
Shown in, it meets the Gauss distribution that peak value is 0.The kurtosis of stochastic variable fully reflects its distribution feelings compared with normal distribution
Steep under condition, the motion vector that therefore normal motion vector crosses information than steganography has bigger kurtosis value.Motion vector
First-order difference is defined as follows:
The distribution situation of first-order difference before and after embedding according to information, the first-order difference distortion factor of definition vector is as follows:
▽ηiWith SiSeparate, so XiProbability mass function be ηiAnd SiThe volume of probability mass function
Long-pending, use hx[n],hs[n],hη[n] represents X respectivelyi, the probability mass function of S, η, then hx[n]=hs[n]*hη[n]。
Convert it to frequency domain and obtain Hx[m]=Hs[m]Hη[m], H is the discrete Fourier transform of h.HηThe derivation of [m] is as follows:
H can be obtainedη[m]≤hη[0]+2hη[k]=hη[0]+hη[k]+hη[-k]=1
So Hx[n]≤Hs[n]。
Introduce center mass feature, reflect the conversion of frequency domain vector:
Calculate the average of single order vector difference, variance simultaneously and reflect the situation of change of vector correlation
Wherein, Si=Si-Si+1, SiIt is the value of motion vector in i-th block, μ (Si) it is SiAverage, σ4(▽
Si) it is SiVariance square.H [n] is SiProbability mass function, H is the discrete Fourier transform (DFT) of h, μsFor motion vector
Average, q is the amplitude range of motion vector first-order difference.
Motion vector horizontal and the E of vertical component will be calculated,As first-order difference feature.
(2) second differnce feature and third order difference feature:
The embedding of information destroys the original dependency relation of motion vector, and single order feature under normal conditions can be preferable
This distortion of reflection vector, the motion vector of carrier video follows the super-Gaussian distribution figure of 0 peak value, such as 4a, and concealed video
Motion vector follow the Gauss distribution of 0 peak value, such as Fig. 4 b, this is single order feature effectively basic place.For vowing in motion
For being directly embedded into the algorithm of information in amount, first-order difference feature has preferable verification and measurement ratio, but first-order difference calculates simple,
Information makes the distribution of motion vector first-order difference keep constant the easiest when embedding.As followed different points for two classes
In cloth first-order difference, topmost difference is wherein h [-2], h [-1], h [0], h [1], numerical value in h [2], secret
Quantity in the quantity normal video to be less than of h [0] in video motion vector first-order difference, and h [-2], h [-1], h [1], h
[2] quantity in unnecessary normal video is then wanted.Remaining hiI=± 3 ... ± n then difference is little.If the energy when Information hiding
Enough keeping h [-2], h [-1], h [0], h [1], h [2] numerical value basically identical, then first-order difference feature will lose efficacy.In order to overcome
This drawback, keeps constant steganographic algorithm to have higher verification and measurement ratio equally to having single order feature, uses high rank difference
Statistic is as one of feature.
The second differnce of definition adjacent motion vectors:
▽2Si=Si-▽Si+1
Before and after then information embeds, the horizontal and vertical component second differnce of vector can be expressed as follows:
WithRepresent respectively2Xi,▽2S,▽2The probability mass function of η, then due to separate
Property, similar with single order situation, obtainIt is converted to frequency domain haveRoot
According to the derivation of document [12], the center mass function of second differnce meets following relation:
H2[n] is the Fourier's series of second differnce probability mass function.
Calculate the variance of second differnce simultaneously:
So that detection algorithm has higher versatility, the steganographic algorithm with feature retention performance is had more preferably
Detection results, in detection feature, add again the third order difference statistical nature of motion vector, third order difference is defined as follows:
Constant in view of the statistical property such as first-order difference to be kept, then need the motion vector changed more, three jumps
The distribution curve of score value will change, and kurtosis is the finger of the high and steep degree of point of the intensity or distribution curve for weighing distribution
Mark, therefore uses second differnce and third order difference kurtosis value as a part for feature, the kurtosis definition of second order and third order difference
As follows:
WithRepresent motion vector second differnce and the average of third order difference respectively,
WithRepresenting motion vector second differnce and the variance square of third order difference respectively, N is the number of block in frame of video.
By the second differnce statistic of motion vector horizontal component and vertical component
And third order difference statisticAs higher difference feature;
Extraction motion vector first-order difference statistical nature and higher difference statistical nature collectively constitute the detection of steganalysis
Feature.
3) integrated classifier design:
The target of steganalysis is detection existence of secret information in digital multimedia.Digital Media, such as image,
Video and audio frequency, be preferable carrier for steganography method, this is because they comprise the most independent element, these yuan
Element can be modified slightly thus be embedded a classified information.Up to the present the method utilizing statistics to describe son is difficult to these
Carrier accurately models, and this considerably increases the difficulty of steganalysis.It is based particularly in carrier and steganography carrier the system extracted
Meter characteristic estimates that the detection method of potential probability distribution is infeasible.Therefore, detection is typically regarded as an engineering
Supervised classification problem in habit.
Despite the presence of substantial amounts of machine learning method, support vector machine is the most most popular method.This mainly by
In, SVMs has solid Fundamentals of Mathematics, and it is to overcome the most again study based on Statistical Learning Theory and worked as intrinsic dimensionality
Remain to provide good result when of bigger than number of samples.But, the complexity of SVM training reduces its use value, very
To being not on the biggest data set at one, training is also required to for a long time, and this is owing to calculating the Gram matrix representing core
Complexity be proportional to intrinsic dimensionality and training set size product square.And, training process itself is at least training sample
A double optimization problem.These problems solve the scale of problem when limiting SVM application in reality, require the most again to analyze
Person has to build more accurate feature and meets the constraint owing to calculating these complexity that resource causes.Integrated classifier provides
The biggest carries out steganalysis from origin, can not be limited by intrinsic dimensionality and designs some features, can not examine simultaneously
The restriction considering training sample number carries out the learning process of a more quick detector.
Integrated classifier is obtained, often by multiple base learners stand-alone training in one group of carrier frame of video and steganography frame of video
One base learner is exactly a simple grader, and the son at the feature space that (uniformly) is chosen at random set up by this grader
Spatially.
Each base learner Bl, l=1 ..., L, is a Rd→ mapping of 0,1}, wherein 0 expression carrier frame, 1
Represent steganography frame.Although it should be noted that learner is defined on the most full dimensional characteristics RdOn, but all of base learner
The dimension d of feature spacesubCan select the value more much smaller than full dimension d, this makes it possible to be effectively reduced computation complexity.
Although the classification performance of each independent base learner is the most weak, but when the value of L is sufficiently large, after carrying out strategy fusion, accurately
Degree will be greatly improved, and may finally restrain.The decision-making value of each base learner is adjusted to waiting priori
In the case of, the error rate minimizing training is:
Wherein PFA,PMDIt is the probability of false-alarm and missing inspection respectively.The present invention
In, although intrinsic dimensionality is the highest, but the data volume detected for video is very big, and this integrated classifier can greatly
Shorten the detection time and improve classification accuracy.
The present invention uses Fisher linear discriminant to be used as the learning tool of each base learner, and this is mainly due to it
There is relatively low training complexity.The maximum situation of event consumption is the inverse of covariance matrix and their sum in solving class
Time.It addition, such weak and instability grader adds the degree of scatter of integrated study.
See Fig. 3, owing to FLD is the classification tool of a standard, part relevant to Ensemble classifier described in the present invention.
The lth base learner is in training setUpper training obtains, wherein
It is the subset randomly choosed, simultaneouslyBe from set 1 ..., NtrnThe bootstrapping sample that in }, sampling obtains,Often
One base learner can represent by following characteristic vector:
WhereinIt is the average of each class:
It is the Scatter Matrix in class,
Wherein λ is that steadiness parameter is to ensure SW+ λ I is positive definite, thus avoids in practice, works as SWNumber when being unusual or ill
The unstability that value calculates.To a testing feature vector y ∈ Ytst, the l base learner is mapped by calculatingAnd with
Threshold value (this threshold value needs to pre-adjust thus meets intended performance standard) compares thus obtains this base learner
Classification results.After obtaining all L decision-makings, the output of final integrated classifier is by (many without weights of this L decision-making utilization
Number) temporal voting strategy is combined, will all single base learners the summation of booking result and with decision-making value L/2 carry out right
Ratio, provides final decision result.Notice that this threshold value can be adjusted between [0, L], thus it is wrong to control different two classes
Significance level or one complete reception operating characteristic ROC curve of acquisition by mistake.In the grader of this Project design we
It is L/2 by adjusting thresholds, has made false alarm rate and loss meansigma methods minimize.
FLD integrated classifier in units of frame, obtains integrated classifier parameter when training.Owing to video sequence has very
Strong seriality, it will be assumed that minimum embedding unit is not less than 50 frames (about 2 seconds), carries out in units of 50 frame video segments
Detection, statistics is wherein judged as steganography and the number of normal video frame, and by priori error rate to judging to revise, final basis
Fragment is judged to steganography frame of video and is judged to that the quantity of normal video frame is to judge whole fragment.
The motion vector of the present invention accurately extracts, and simplifies FFmpeg and revises, and only retains and therein h.264 decodes
Part, and increase the output function of motion vector information, while decoding, realize the extraction of motion vector quickly and accurately,
Highly versatile.Vector characteristic extraction algorithm devises higher difference feature on the basis of first-order difference feature, this category feature energy
The steganographic algorithm having vector statistics feature to compensate preferably carries out steganalysis, and detection efficiency is high, versatility, stability
By force.Video image characteristic quantity is big, processes by single complex classifier, such as support vector machine or Method Using Relevance Vector Machine
The required detection time is the longest, and with simple classification device such as FLD, then Detection results is poor, time efficiency and accuracy rate it
Between be difficult to seek an equilibrium point, the present invention integrated steganography classifier design utilizes the thought of integrated study, to multiple simply
Grader is trained respectively, designs a kind of integrated steganography grader based on FLD.Integrated classifier is not only on classification speed
It is greatly improved, and than using single same category utensil to have higher classification accuracy.Design resisted motion vector
The steganalysis algorithm that information embeds, and realize its software system.On the basis of feature extraction and classifier design, it is designed to
The steganalysis algorithm that opposing motion vector information embeds, and with H.264 carrier for detection object, it is achieved steganalysis software
System.The video steganalysis system that the present invention is developed, it is achieved function be effectively analyze judge based on H.264/
The video that AVC standard motion vector information embeds, verification and measurement ratio reaches more than 70%.
Claims (10)
1. the steganalysis algorithm embedded for H.264 video motion vector information, it is characterised in that include following step
Rapid:
1) motion vector extracts: is decoded video, obtains the motion vector of each macro block, as motion vector Differential Characteristics
The data extracted;
2) to step 1) data that obtain carry out motion vector Differential Characteristics extraction:
2.1) first-order difference feature extraction:
First-order difference distortion factor E of calculating motion vector:
Wherein,It is the kurtosis value of motion vector first-order difference,For motion vector first-order difference,SiBeing the value of motion vector in i-th block, h [-2], h [-1], h [0], h [1], h [2] represent respectively's
Probability mass function;
Calculate the average of single order vector difference
Wherein, q is the amplitude range of motion vector first-order difference, hS[n] isProbability mass function;
Calculate the variance of single order vector difference
Wherein, μsFor the average of motion vector, hS[n] isProbability mass function;
By E,As first-order difference feature;
2.2) second differnce feature and third order difference feature extraction:
The center mass function of calculating motion vector second differnce:
Wherein,Represent respectivelySecond differnceThe Fourier of probability mass function
Leaf changes, and i is the number of block in a frame, and N is the number of block in frame of video;
The variance of calculating second differnce:
Wherein, μsFor the average of motion vector,It it is second differnceProbability mass function, q is motion vector one jump
The amplitude range divided;
Calculating second differnce and the kurtosis value of third order difference:
Wherein,WithRepresent motion vector second differnce and the average of third order difference respectively,Representing motion vector second differnce and the variance square of third order difference respectively, N is frame of video
The number of middle piece;
By second differnce statisticAnd third order difference statistic
As higher difference feature;
3) FLD integrated classifier is used according to first-order difference feature and higher difference feature, the video segment of input to be detected
Judging, statistics is wherein judged as that the number of steganography and normal video frame, to judge whole video segment, i.e. completes steganalysis and calculates
Method.
A kind of steganalysis algorithm embedded for H.264 video motion vector information the most according to claim 1, it is special
Levy and be, described step 1) in use FFmpeg decoder video to be decoded, at decode_slice_header () letter
Number obtains the frame number of current decoded frame, h1_motion () function obtains the type of each macro block, at mc_dir_
Part () function obtains the motion vector of each macro block, and exports in TXT text, extract as motion vector Differential Characteristics
Data.
A kind of steganalysis algorithm embedded for H.264 video motion vector information the most according to claim 1, it is special
Levy and be, described step 2.1) in the computing formula of kurtosis value of motion vector first-order difference as follows:
Wherein,It is the average of motion vector first-order difference,Represent that the variance of motion vector first-order difference is put down
Side, N is the number of block in frame of video.
A kind of steganalysis algorithm embedded for H.264 video motion vector information the most according to claim 3, it is special
Levy and be, described step 2.1) in introduce the center mass feature of first-order difference function and reflect motion vector first-order difference
Energy distribution situation, by center mass feature C after calculating first-order difference probability mass function Fourier transform1(H [m]),
Reflect that vector first-order difference is in frequency domain energy variation situation:
Wherein, H [i] represents, T represents, i represents the number of block in a frame.
A kind of steganalysis algorithm embedded for H.264 video motion vector information the most according to claim 4, it is special
Levy and be, described step 2.1) in the first-order difference of motion vector be defined as follows:Carrier video is transported
The first-order difference Distribution value of dynamic vector meets the super-Gaussian distribution that peak value is 0, the horizontal component after embedding information and vertical component
Motion vector first-order differenceWithIt is expressed as:
Wherein,WithIt is respectively horizontal component and the embedding information of vertical component of first-order difference,WithIt is respectively water
The amount of dividing equally and the first-order difference of vertical component.
A kind of steganalysis algorithm embedded for H.264 video motion vector information the most according to claim 1, it is special
Levy and be, described step 2.2) defined in the second differnce of adjacent motion vectors be:
After information embedding, the horizontal and vertical component second differnce of motion vector is expressed as:
Wherein,WithIt is respectively horizontal component and the embedding information of vertical component of second differnce,WithRespectively
For horizontal component and the first-order difference of vertical component.
A kind of steganalysis algorithm embedded for H.264 video motion vector information the most according to claim 1, it is special
Levy and be, described step 2.2) in third order difference be defined as follows:
A kind of steganalysis algorithm embedded for H.264 video motion vector information the most according to claim 1, it is special
Levy and be, described step 3) in FLD integrated classifier include several base learners, each base learner Bl, l=
1 ..., L, is Rd→ { mapping of 0,1}, wherein 0 represents carrier frame, and 1 represents steganography frame, RdFor all full dimensional characteristics, often
The decision-making value of one base learner is adjusted in the case of waiting priori, and the error rate minimizing training is:
Wherein PFA,PMDIt is the probability of false-alarm and missing inspection respectively.
A kind of steganalysis algorithm embedded for H.264 video motion vector information the most according to claim 8, it is special
Levying and be, each described base learner uses Fisher linear discriminant to be used as learning tool.
A kind of steganalysis algorithm embedded for H.264 video motion vector information the most according to claim 9, its
Being characterised by, the lth base learner of described FLD integrated classifier is in training setUpper training
Obtain, whereinIt is the subset randomly choosed, simultaneouslyBe from set 1 ..., NtrnAdopt in }
The bootstrapping sample that sample obtains,The characteristic vector of each base learner is expressed as:
WhereinIt is the average of each class: Being the Scatter Matrix in class, wherein λ is steady
Determine parameter to ensure SW+ λ I is positive definite.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610177535.5A CN105872555B (en) | 2016-03-25 | 2016-03-25 | A kind of steganalysis algorithm for the insertion of H.264 video motion vector information |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610177535.5A CN105872555B (en) | 2016-03-25 | 2016-03-25 | A kind of steganalysis algorithm for the insertion of H.264 video motion vector information |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105872555A true CN105872555A (en) | 2016-08-17 |
CN105872555B CN105872555B (en) | 2019-01-15 |
Family
ID=56625225
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610177535.5A Expired - Fee Related CN105872555B (en) | 2016-03-25 | 2016-03-25 | A kind of steganalysis algorithm for the insertion of H.264 video motion vector information |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105872555B (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107172434A (en) * | 2017-06-27 | 2017-09-15 | 郑州师范学院 | A kind of method that built-in steganography is carried out to H.265 video |
CN107197298A (en) * | 2017-06-27 | 2017-09-22 | 郑州师范学院 | A kind of reversible steganography method of HEVC videos of the undistorted drift based on multi-component system |
CN107333139A (en) * | 2017-06-27 | 2017-11-07 | 郑州师范学院 | A kind of H.265 half invertible video steganography information approach based on intra prediction mode |
CN107947898A (en) * | 2017-11-15 | 2018-04-20 | 深圳大学 | Information detecting method, device and receiving device based on optimization packet variance |
CN108460405A (en) * | 2018-02-02 | 2018-08-28 | 上海大学 | A kind of image latent writing analysis Ensemble classifier optimization method based on deeply study |
CN109361934A (en) * | 2018-11-30 | 2019-02-19 | 腾讯科技(深圳)有限公司 | Image processing method, device, equipment and storage medium |
CN110062242A (en) * | 2019-03-04 | 2019-07-26 | 中山大学 | A kind of H.264 video steganographic algorithm based on UED |
CN110149457A (en) * | 2019-05-08 | 2019-08-20 | 中山大学 | A kind of multiple-objection optimization of with constraint conditions H.264 video steganography method |
CN110324634A (en) * | 2019-07-05 | 2019-10-11 | 中国科学技术大学 | It is a kind of to be embedded in the video steganography method that distortion is decomposed based on motion vector |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR100734845B1 (en) * | 2005-11-18 | 2007-07-03 | 한국전자통신연구원 | Method for discriminating malicious video using image quality classification and apparatus thereof |
CN103338376A (en) * | 2013-07-02 | 2013-10-02 | 中国科学技术大学 | Video steganography method based on motion vector |
CN103888773A (en) * | 2014-02-19 | 2014-06-25 | 南京邮电大学 | Video steganography analysis method based on mutual information and motion vectors |
CN104837011A (en) * | 2015-05-04 | 2015-08-12 | 中国科学院信息工程研究所 | Content self-adaptive video steganalysis method |
-
2016
- 2016-03-25 CN CN201610177535.5A patent/CN105872555B/en not_active Expired - Fee Related
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR100734845B1 (en) * | 2005-11-18 | 2007-07-03 | 한국전자통신연구원 | Method for discriminating malicious video using image quality classification and apparatus thereof |
CN103338376A (en) * | 2013-07-02 | 2013-10-02 | 中国科学技术大学 | Video steganography method based on motion vector |
CN103888773A (en) * | 2014-02-19 | 2014-06-25 | 南京邮电大学 | Video steganography analysis method based on mutual information and motion vectors |
CN104837011A (en) * | 2015-05-04 | 2015-08-12 | 中国科学院信息工程研究所 | Content self-adaptive video steganalysis method |
Non-Patent Citations (2)
Title |
---|
张英男 等: "基于灰色关联分析的H_264_AVC视频隐写算法", 《武汉大学学报(理学版)》 * |
王珏 等: "基于运动矢量分量差的视频隐写算法", 《计算机工程》 * |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107197298A (en) * | 2017-06-27 | 2017-09-22 | 郑州师范学院 | A kind of reversible steganography method of HEVC videos of the undistorted drift based on multi-component system |
CN107333139A (en) * | 2017-06-27 | 2017-11-07 | 郑州师范学院 | A kind of H.265 half invertible video steganography information approach based on intra prediction mode |
CN107172434B (en) * | 2017-06-27 | 2019-06-18 | 郑州师范学院 | A method of built-in steganography is carried out to H.265 video |
CN107172434A (en) * | 2017-06-27 | 2017-09-15 | 郑州师范学院 | A kind of method that built-in steganography is carried out to H.265 video |
CN107947898A (en) * | 2017-11-15 | 2018-04-20 | 深圳大学 | Information detecting method, device and receiving device based on optimization packet variance |
CN107947898B (en) * | 2017-11-15 | 2020-08-07 | 深圳大学 | Information detection method and device based on optimized grouping variance and receiving equipment |
CN108460405A (en) * | 2018-02-02 | 2018-08-28 | 上海大学 | A kind of image latent writing analysis Ensemble classifier optimization method based on deeply study |
CN109361934B (en) * | 2018-11-30 | 2021-10-08 | 腾讯科技(深圳)有限公司 | Image processing method, device, equipment and storage medium |
CN109361934A (en) * | 2018-11-30 | 2019-02-19 | 腾讯科技(深圳)有限公司 | Image processing method, device, equipment and storage medium |
US11798145B2 (en) | 2018-11-30 | 2023-10-24 | Tencent Technology (Shenzhen) Company Limited | Image processing method and apparatus, device, and storage medium |
CN110062242A (en) * | 2019-03-04 | 2019-07-26 | 中山大学 | A kind of H.264 video steganographic algorithm based on UED |
CN110062242B (en) * | 2019-03-04 | 2021-05-04 | 中山大学 | H.264 video steganography algorithm based on UED |
CN110149457A (en) * | 2019-05-08 | 2019-08-20 | 中山大学 | A kind of multiple-objection optimization of with constraint conditions H.264 video steganography method |
CN110149457B (en) * | 2019-05-08 | 2021-05-04 | 中山大学 | Multi-objective optimization H.264 video steganography method with constraint conditions |
CN110324634B (en) * | 2019-07-05 | 2021-10-01 | 中国科学技术大学 | Video steganography method based on motion vector embedding distortion decomposition |
CN110324634A (en) * | 2019-07-05 | 2019-10-11 | 中国科学技术大学 | It is a kind of to be embedded in the video steganography method that distortion is decomposed based on motion vector |
Also Published As
Publication number | Publication date |
---|---|
CN105872555B (en) | 2019-01-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105872555A (en) | Steganalysis algorithm specific to H.264 video motion vector information embedment | |
CN102917227B (en) | Compressive sensing-based adaptive video information hiding method | |
CN104519361A (en) | Video steganography analysis method based on space-time domain local binary pattern | |
CN105844238A (en) | Method and system for discriminating videos | |
CN103679719A (en) | Image segmentation method | |
CN104200203A (en) | Human movement detection method based on movement dictionary learning | |
CN104282310A (en) | Steganography detection method for audio subjected to MP3Stego steganography | |
CN103714340B (en) | Self-adaptation feature extracting method based on image partitioning | |
CN107292315B (en) | Steganalysis method and steganalysis device based on multi-scale LTP (low temperature processing) features | |
CN107027051A (en) | A kind of video key frame extracting method based on linear dynamic system | |
CN113283577A (en) | Industrial parallel data generation method based on meta-learning and generation countermeasure network | |
Wang et al. | A posterior evaluation algorithm of steganalysis accuracy inspired by residual co-occurrence probability | |
CN116012835A (en) | Two-stage scene text erasing method based on text segmentation | |
CN103903271B (en) | Image forensics method for natural image and compressed and tampered image based on DWT | |
Mooers et al. | Generative modeling of atmospheric convection | |
CN105913008B (en) | Based on the assumption that the crowd's accident detection method examined | |
Zhang et al. | Coverless video steganography based on audio and frame features | |
Ren et al. | Student behavior detection based on YOLOv4-Bi | |
CN115563510B (en) | Training method of click rate estimation model and related device | |
CN110554429A (en) | Earthquake fault identification method based on variable neighborhood sliding window machine learning | |
CN109192217A (en) | General information towards multiclass low rate compression voice steganography hides detection method | |
CN113486932B (en) | Convolutional neural network steganalysis-oriented optimization method | |
CN106303696B (en) | H.264 video content authenticating water-mark method based on pivot information distribution proportion stability | |
CN106791871B (en) | A kind of hiding detection method of motion vector modulation intelligence | |
Niu et al. | An image steganography method based on texture perception |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190115 |
|
CF01 | Termination of patent right due to non-payment of annual fee |