CN106228505B - A kind of robust general steganalysis method of picture material perception - Google Patents
A kind of robust general steganalysis method of picture material perception Download PDFInfo
- Publication number
- CN106228505B CN106228505B CN201610567405.2A CN201610567405A CN106228505B CN 106228505 B CN106228505 B CN 106228505B CN 201610567405 A CN201610567405 A CN 201610567405A CN 106228505 B CN106228505 B CN 106228505B
- Authority
- CN
- China
- Prior art keywords
- image
- dictionary
- block
- image block
- follows
- 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.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T1/00—General purpose image data processing
- G06T1/0021—Image watermarking
- G06T1/005—Robust watermarking, e.g. average attack or collusion attack resistant
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2201/00—General purpose image data processing
- G06T2201/005—Image watermarking
- G06T2201/0202—Image watermarking whereby the quality of watermarked images is measured; Measuring quality or performance of watermarking methods; Balancing between quality and robustness
Abstract
The purpose of the present invention is to provide a kind of picture material perception robust general steganalysis method, method of the invention can preferably utilize CS technical treatment digital picture the advantages of, further increase steganalysis data-handling efficiency and classification judge precision.Specifically comprise the following steps: the size N, piecemeal size B and measured value size M of (1) setting digital picture X；(2) 9/7-Haar DWT sparse matrix Ψ is determined；(3) the Hadamard calculation matrix Φ of BCS is determinedB；(4) digital picture is divided into B × B block, and according to Hadamard calculation matrix ΦB；Calculate the measurement vector y of i-th of image blockiWith the measurement data Y of whole image X；(5) smooth and non-smooth perceptually result are divided the image into；(6) training process；(7) assorting process；(8) judgement, analytic process；(9) iteration is searched for, until detecting complete image data.
Description
Technical field
The present invention relates to the network information securitys, relate in particular to a kind of robust general steganalysis of picture material perception
Method.
Background technique
Steganalysis is the important component of the network information security, its most root problem is to judge digital carrier
Whether secret information is carried.Steganalysis is the countermeasure techniques of Information hiding, is the data Transformation Attack to Steganography, it is therefore an objective to
In order to detect the presence of classified information in Steganography, secret information is extracted to identify and destroy secret communication.But it is existing general
Blind checking method does not account for influence of the picture material to steganalysis performance mostly, and does not account for hidden under Prerequisite
Information blind Detecting mode is hidden, for this purpose, the technical problem to be solved by the present invention is
(1) it is directed to the different images content (such as texture and smooth) in airspace, proposes a kind of new compressed sensing
The general steganalysis method in the domain CS (Compressive Sensing, CS).This method by the perception to picture material at
Reason, solves the structural classification problem of digital picture different content.
(2) redundant dictionary of invention different images content, and by the dictionary updating to picture material it is smooth, sharpen,
The attacks such as diminution, shearing and recompression are handled, and solve the problems, such as the robust detection of image general steganalysis.
Steganalysis mainly includes two aspects: dedicated steganalysis and general steganalysis.Dedicated steganalysis is directed to
The analysis method that known steganographic algorithm proposes, it, which needs to determine first just make in stego-image using which kind of Steganography, sentences
Disconnected, accuracy rate is higher, but restricted application.General steganalysis is on the basis of unknown initial carrier object and Steganography
To detection image whether containing a kind of close analysis method judged.Since universal blind checking method does not need related steganographic algorithm
The prior information of details, and with the raising of feature extraction validity and classifier ability, the detection of universal blind checking method
Performance is gradually increased, and has become the mainstream of steganalysis research.Universal blind checking method is mainly used based on machine learning
Method, key is the feature found energy effective district fractional bearer and carry close image, therefore the difference of each method lies also in and extracted
Characteristic of division it is different.
But since in the covert communications detection system under real network environment, the video source of acquisition usually contains a variety of more
The picture material of sample, this has apparent difference with the steganalysis algorithm research under laboratory environment compared with single image content,
General blind Detecting classifier be will cause in " mismatch " problem of training stage and the carrier image source statistics of test phase,
To reduce the detection effect of detection algorithm, this brings tired to the steganographic detection under network isomery image-context.
Therefore, currently, the robust general steganalysis method based on digital image content perception at home and abroad or blank.
Similar method is [AMIRKHANI H, the RAHMATI M.New framework for using image such as Amirkhani
Contents in blind steganalysis systems [J] .Journal of Electronic Imaging, 2011,
20 (1): 1-14.] for a kind of general blind Detecting frame of different images content proposition；Another method is Hashemipour
Deng [HASHEMIPOUR E M, RAHMATI M.A statistical blind image steganalysis based on
image multi-classification[C]//Proceedings of the 2012Eighth International
Conference on Intelligent Information Hiding and Multimedia Signal
Processing.Piscataway, NJ:IEEE Press, 2012:150-153.] propose one kind be based on image multi-categorizer
Blind Detecting frame.Fig. 1 (a) and 1 (b) is the synthesis realization principle figure of both methods.Both methods is limited in that:
(1) steganalysis of general blind Detecting frame: the mentioned classification method of this method assumes carrier and carries close figure accordingly
As complexity is equal.But in fact, the insertion of classified information changes the correlation between image adjacent pixel, therefore also change
The complexity of image, carrier and carries close image accordingly and is possible to not in same class；And only with the classification method averagely classified
It is simply to carry out average classification according to the range of image complexity and the classification to be divided into, it cannot be guaranteed obtained by each classification
The detection performance arrived is optimal.
(2) steganalysis of multi-categorizer blind Detecting frame: this method is exchanged just with jpeg image non-zero
(Alternating Current, AC) coefficient ratio as the more characteristic of division of image, cannot picture engraving content well it is more
Sample less can determine that the position carried in close image where secret information.
Compressive sensing theory (Compressed Sensing, CS) is emerging important in signal processing and art of mathematics
Theory, essence are can to restore entire signal by a small amount of measurement data with high probability to sparse or compressible signal (such as image).
Its main thought is the immanent structure for using complete dictionary that traditional orthogonal basis function is replaced to disclose signal, and can be by dilute
It dredges expression, incoherent measurement and restructing algorithm (presently mainly convex optimized algorithm and greedy algorithm) and restores original signal.
Since the general steganalysis of digital image content focuses on structural analysis and feature extraction, and CS technology has
Have that sampled data output is few, structure feature is obvious and high probability restores the characteristic of signal, so, it is suitble to believe image under highly concealed type
Number data processing, secret information extract requirement.
Based on this, the present invention proposes " in a kind of image from the angle of the detection content of steganalysis and robustness
Hold the robust general steganalysis method of perception ", from the basic of the data processing of steganalysis, dictionary learning and content analysis
A kind of new invention achievement is explored in problem for information security.
Summary of the invention
The purpose of the present invention is to provide a kind of robust general steganalysis method of picture material perception, sides of the invention
Method can preferably utilize the advantages of CS technical treatment digital picture, further increase the data-handling efficiency of steganalysis and divide
Class judges precision.
The purpose of the present invention can be realized by technical measures below:
(1) the size N, piecemeal size B and measured value size M of digital picture X are set；
(2) 9/7-Haar DWT sparse matrix Ψ is determined:
(3) the Hadamard calculation matrix Φ of BCS is determinedB:
(4) digital picture is divided into B × B block, and according to Hadamard calculation matrix ΦBCalculate the measurement of i-th of image block
Vector yiWith the measurement data Y of whole image X:
Wherein, yi=ΦB·xi, i=1,2 ..., B2；
The then measurement data Y of whole image X are as follows:
(5) according to the energy of each image block of BCS measured valueI=1,2 ..., B × B are calculatedAnd judge the measurement vector y of i-th of image blockiWhether threshold value T=is more than or less than
0.4E, to divide the image into smooth and non-smooth perceptually result；
(6) training process；
(7) assorting process；
(8) judgement, analytic process；
(9) iteration is searched for, until detecting complete image data.
Specific step is as follows for step (6) training process:
(6.1) according to the different digital perception of content of step (5) as a result, determining RDWT dictionary and orthogonal DCT dictionary: being formed
Dictionary D=[D1, D2 ..., D13]；
2 grades of RDWT dictionary formats in dictionary D are as follows:
Wherein, symbolIt indicates inner product operation, is a kind of operation of element multiplication between matrix；
The form of orthogonal DCT dictionary are as follows:
(6.2) the sparse initial value of dictionary is calculatedAnd calculate image block xiInitial value:
Described image block xiThe specific calculating process of initial value is as follows: setting diFor middle dictionary DiAtom, RiFor iteration ginseng
Difference: Ri=| | ΦBxi-ΦBxi-1||2=| | yi-ΦBxi-1||2,The sparse initial value of dictionaryIt is dictionary original
Sub- diWith the irregular R of iterationiInner product, the subscript 0 of symbol indicates the 0th value, that is, initial value；
It obtains in this way, image block xiInitial valueForm are as follows:
Wherein, n is data vector length and dictionary size, i=1,2 ..., n；The subscript 0 of symbol indicates the 0th value,
It is exactly initial value；
(6.3) image block x is calculatedi=Diαi, and according to formula: Ri=| | ΦBxi-ΦBxi-1||2=| | yi-ΦBxi-1||2
Repetitive exercise updates dictionary D；
Calculating process similar step (6.2), way of realization is as follows:
Specific step is as follows for step (7) assorting process:
(7.1) according to formula yi=ΦB·xi, calculate the measured value y of i-th of image blockiAnd its energyAnd combine the relationship of energy and threshold valueDetermine digital picture not
With the data area of content；
(7.2) energy of each image block is calculatedSize: the respectively ceiling capacity of image and minimum energy are set
Amount, calculates the testing time of imageEmFor image energy mean value；
(7.3) according to formulaDetermine the measured value y of i-th of image blocki
In the sparse coefficient of j-th of dictionaryWherein j=1,2 ..., 13；
Way of realization it is as follows:
(7.4) it iterates to calculate, obtaining image sparse indicates coefficientWherein i=1,2 ...,
3136；According to the different content of the non-smooth block peace sliding block of digital picture, according to being ordered from large to small image sparse table
Show factor alpha.
Specific step is as follows for the step (8):
(8.1) according to formula yi=ΦB·xiWithCalculate separately the measurement vector y of i-th of image blockiWith
The measurement data Y of whole image X；According to formulaThe energy of i-th of image block is calculated,
And combine the relationship of its energy size and threshold valueDetermine the data area of i-th of image block: if Ei≤
T then belongs to the smooth block of digital picture；If Ei> T then belongs to the non-smooth block of digital picture；
The calculating process of the measurement data Y of (8.1.1) image X is as follows:
The measurement vector y of i-th of image blockiAre as follows: yi=ΦB·xi, wherein i=1,2 ..., B2；
The then measurement data Y of whole image X are as follows:
(8.1.2) determines that the non-smooth block peace sliding block process of digital picture is as follows:
According to BCS measured value yi, calculate the energy of i-th of image blockWherein i
=1,2 ..., B × B (B is piecemeal size), and according toDivide the image into smooth and non-partes glabra
Point；
(8.2) searching and xiSimilar L image block obtains being indicated with the linear combination of L similar block to present image
Block xiEstimated value xi′:
Wherein,Indicating the design factor of similar block, its value is directly proportional to the noise variance of image in the area,
The noise variance is estimated with steady median operator；
(8.3) Hadamard calculation matrix ΦBSubstitute into formula:
Wherein β is iterative parameter, and value range is | Ri-Ri| < β, 0 < β < 1 update the piecemeal of current survey image
Calculated valueK is iteration optimization number；
(8.4) according to formulaIt is solved with Lasso algorithm
To αiInitial solutionAnd substitute into formulaUpdate image block redundant dictionary
Coefficient, and substitute into following formula:
Seek the sparse bayesian learning coefficient of detection image block
Wherein, ΦBFor piecemeal Hadamard calculation matrix, y is measured value, and d is the atom in dictionary D, and R is iterative calculation
Error, α are dictionary coefficient, i=1,2 ..., n, n=3136；
(8.5) analysis judgement output: ifThen i-th of image block contains secret data in jth block dictionary,
Secret data position is xi；IfThen i-th of image block is free of secret data in jth block dictionary, is clean original
Beginning image block.
The advantages of general steganalysis method proposed by the present invention that digital content selection and classification are carried out based on CS technology
Are as follows:
(1) accidental projection measurement is carried out to digital picture using piecemeal Hadamard calculation matrix, and according to measurement data
Image is divided into smooth and non-smooth different piece content by energy statistics distribution.This division can sufficiently illustrate image
Texture structure and smooth structure characteristic information, improve the treatment effeciency of data.
(2) according to the digital content of different piece, the image block in natural image library by its directionality from 0 ° to 180 °
12 class RDWT (redundant discrete small wave converting) dictionaries and a kind of orthogonal DCT (discrete orthogonal transform) dictionary are established, form 3136
The redundant dictionary D of atom.Orthogonal DCT dictionary in the dictionary can indicate the isotropism feature in image block, and RDWT word
Allusion quotation can indicate the angle and direction in image.Dictionary of the invention, which enhances, divides different digital picture material steganalysis
Class adaptability.(3) letter is approached by the similarity measurement of the different digital content to detection image and its dictionary sparse coefficient
Number realizes accurate judgement and position output containing close digital picture.With the similarity measurement and dictionary of digital image content
The approximation technique of sparse coefficient improves the judgement precision of steganalysis, it is determined that the secret location of stego-image.
Detailed description of the invention
Fig. 1 (a) and 1 (b) is the steganalysis implementation schematic diagram of prior art image content-based；
The present invention is based on the general steganalysis method flow charts that picture material perceives by Fig. 2；
Fig. 3 (a) and 3 (b) is the dictionary D according to different images content study construction；
Fig. 4 is to realize the experimental result indicated the dictionary of different images by this method in image library；
Fig. 5 is the experiment for the Secret Image position that can determine to stego-image by this method dictionary learning in image library
Result schematic diagram.
Specific embodiment
The present invention carries out accidental projection to image block using CS piecemeal hadamard matrix, big further according to the energy of observation vector
It is small observed image is divided into non-smooth sliding two parts of peace to be trained classification, detection is finally updated according to classifying dictionary iteration and is schemed
It seem the no position containing secret information and determining secret information.
The robust general steganalysis method of picture material perception of the invention is based on CS technical principle, is divided into digital picture
Content test training and digital image content classification deterministic process, include following content based on CS technical principle:
(1) the compressed sensing piecemeal processing of image:
If data image signal X ∈ RNThe signal that length to obtain from M sampled signal is N, and M < < N.So,
To measuring signal Y:
Y=AX (1)
Wherein, the length of Y is M, and A is M × N calculation matrix (also referred to as observing matrix), and has sub-sampling rate S=M/
N.If X be at some transformation matrix Ψ it is sparse, that is, have:
X=Ψ α, | | α | |0< K (2)
Here, transformation matrix Ψ uses 9/7-Haar DWT sparse matrix, and α is sparse coefficient, | | α | |0Indicate 0- norm,
That is the non-zero number of sparse coefficient.At this moment, (1) formula becomes:
Y=A Ψ α (3)
When measurement dimension meets K < M < N, signal X can be by solving the optimization problem under 0- norm or 1- norm from M
High probability reconstructs in=O (KlogN) a measurement data.
For two dimensional image signal, since the length of N is very big.For this purpose, the present invention uses piecemeal CS technology (Block-
based CS,BCS)[Gan L.Block compressed sensing of natural images[C].In
Proceedings of the International Conference on Digital Signal Processing,
Cardiff, UK, July 2007, pp.403-406.] piecemeal perception processing is carried out to image.In BCS, an image is divided
It is sampled at B × B block, and using the calculation matrix of an approximate size.Assuming that xiIndicate that image X passes through the scanning input of Z row
I-th of image block vector, then having:
yi=ΦB·xi (4)
Wherein, ΦBFor the calculation matrix of B × B size.The measurement data of this sampled images X are as follows:
The sub- rate of the sampling of whole image X is S=MB/B2.At this moment, the calculation matrix A of image X has diagonal form knot in (1) formula
Structure matrix A=diag (Φ), form are as follows:
Calculation matrix A realizes that image data is seen using piecemeal hadamard matrix BHM (Block Hadamard Matrix)
It surveys.The structural form of n rank BHM is as follows:
As image block size B=2nWhen, the calculation matrix A of formula (6) are as follows:
Then hadamard matrix ΦBWay of realization are as follows:
For extracted with high accuracy image data, according to BCS value yiEnergy
BCS measurement data is divided into smooth and texture part according to (10) formula by size and its threshold value T of determination:
Wherein, T=0.4Em, andFor the mean value of observation energy.
(2) dictionary learning of image block
According to CS technology, since redundant dictionary can indicate image in more sparse mode, so, rarefaction representation is by signal
Energy concentrates on a small amount of atom, these atoms contain the key structural feature of image.Therefore, in conjunction with smoothed image block and non-
The structure feature of smoothed image block is different, and the present invention selects different classifying dictionaries to smooth block and non-smooth block.Wherein, non-flat
Cunning image block selection redundancy (Redundant Discrete Wavelet Transform, RDWT) dictionary [Fowler J.E.,
“The redundant discrete wavelet transform and additive noise,”[J],IEEE Signal
Processing, 2005,12 (9): 629-632.], and the structure feature of smoothed image block is fairly simple, selects orthogonal DCT word
Allusion quotation.
Image block in natural image library ° is divided into 12 classes from 0 ° to 180 by its directionality, every class data use K- respectively
SVD (K-Singular Value Decomposition) method trains a category dictionary [M Aharon, M Elad, A
Bruckstein.K-SVD:An algorithm for designing over complete dictionaries for
sparse representation[J].IEEE Transaction.on Signal Processing,2006,54(11):
4311-4322.], every category dictionary includes 256 atoms, in addition along with orthogonal DCT dictionary is formed including 3136 atoms altogether
Redundant dictionary D.Orthogonal DCT dictionary can indicate the isotropism feature in image block, and RDWT dictionary can indicate image
In angle and direction.
In CS projection domain, if i-th of image block of image X and the measurement data of (i-1) a image block are respectively xi,
xi-1, RiIt is irregular for iteration, then RiIterative process is as follows:
Ri=| | ΦBxi-ΦBxi-1||2=| | yi-ΦBxi-1||2 (11)
When | Ri-Ri| when < β, 0 < β < 1, dictionary D iteration stopping.Fig. 3 (a) and 3 (b) is according to different images content
Practise the dictionary result of the Lena image of construction.
(3) classification of picture material
Design one be adapted with feature, coincide with picture structure, with maximum discrimination score (namely the smallest false-alarm
Rate and false dismissed rate) classifier be also a most important step.
If αiFor based on dictionary DiThe rarefaction representation coefficient vector of ∈ D,For in dictionary DiThe rarefaction representation of middle classification and Detection
Coefficient vector utilizes Lasso (Least Absolute Shrinkage and Selection Operator) algorithm [Dong
Wei-sheng,Zhang Lei,Shi Guang-ming,Wu Xiao-lin.Image deblurring and super-
resolution by adaptive sparse domain selection and adaptive regularization
[J] .IEEE Transactionson Image Processing, 2011,20 (7): 1838-1857.] realize image block in D
In rarefaction representation coefficient such as formula (12) under each dictionary:
D in formulajFor jth in D (j=1 ..., 13) a dictionary, ρ is weight coefficient, ρ=max | (ΦBDj)Tyi|}.It will be to
Rarefaction representation coefficient of the detection image under redundant dictionaryThe processing of l1 norm minimum is carried out, and dictionary iteration minimal error β
It as constraint condition, is added to during classification and Detection, sparse bayesian learning function can be obtained are as follows:
Wherein, λ is weight coefficient, x 'iUtilize the secret information estimated value for the stego-image block that similar image block obtains.Root
X is enabled according to CS technologyi=Diαi, formula (13) are solved by the method that alternating iteration optimizes k solution and are obtained:
(14) first item in formula is to meet observational equation yi=ΦBxi, Section 2 is for determining depositing containing confidential information
In position.(15) first item in formula is in order to closest with the solving result of formula (14), and Section 2 is to meet image block
The most sparse principle of expression coefficient under dictionary.Formula (14) are solved using gradient descent algorithm:
The iteration of formula (17) needs first to determine initial valueSolving following optimization problem with Lasso algorithm can be obtained αiJust
Begin solution:
Image block x is obtained by sparse inverse transformation againiInitial value:
According to principles above, process as shown in Figure 2, the robust general steganalysis method of picture material perception of the invention
Include the following steps:
(1) the size N, piecemeal size B and measured value size M of digital picture X are set；
(2) 9/7-Haar DWT sparse matrix Ψ is determined:
(3) the Hadamard calculation matrix Φ of BCS is determinedB:
(4) digital picture is divided into B × B block, and according to Hadamard calculation matrix ΦBCalculate the measurement of i-th of image block
Vector yiWith the measurement data Y of whole image X:
Wherein, yi=ΦB·xi, i=1,2 ..., B2；
The then measurement data Y of whole image X are as follows:
(5) according to the energy of each image block of BCS measured valueI=1,2 ..., B × B are calculatedAnd judge the measurement vector y of i-th of image blockiWhether threshold value T=is more than or less than
0.4E, to divide the image into smooth and non-smooth perceptually result；
(6) training process:
(6.1) according to the different digital perception of content of step (5) as a result, determining RDWT dictionary and orthogonal DCT dictionary: being formed
Dictionary D=[D1, D2 ..., D13]；
2 grades of RDWT dictionary formats in dictionary D are as follows:
Wherein, symbolIt indicates inner product operation, is a kind of operation of element multiplication between matrix；
The form of orthogonal DCT dictionary are as follows:
(6.2) the sparse initial value of dictionary is calculatedAnd calculate image block xiInitial value:
Described image block xiThe specific calculating process of initial value is as follows: setting diFor middle dictionary DiAtom, RiFor iteration ginseng
Difference: Ri=| | ΦBxi-ΦBxi-1||2=| | yi-ΦBxi-1||2,The sparse initial value of dictionaryIt is dictionary original
Sub- diWith the irregular R of iterationiInner product, the subscript 0 of symbol indicates the 0th value, that is, initial value；
It obtains in this way, image block xiInitial valueForm are as follows:
Wherein, n is data vector length and dictionary size, i=1,2 ..., n；The subscript 0 of symbol indicates the 0th value,
It is exactly initial value；
(6.3) image block x is calculatedi=Diαi, and according to formula: Ri=| | ΦBxi-ΦBxi-1||2=| | yi-ΦBxi-1||2
Repetitive exercise updates dictionary D；
Calculating process similar step (6.2), way of realization is as follows:
(7) assorting process:
(7.1) according to formula yi=ΦB·xi, calculate the measured value y of i-th of image blockiAnd its energyAnd combine the relationship of energy and threshold valueDetermine digital picture not
With the data area of content；
(7.2) energy of each image block is calculatedSize: the respectively ceiling capacity of image and minimum energy are set
Amount, calculates the testing time of imageEmFor image energy mean value；
(7.3) according to formulaDetermine the measured value y of i-th of image blocki
In the sparse coefficient of j-th of dictionaryWherein j=1,2 ..., 13；
Way of realization it is as follows:
(7.4) it iterates to calculate, obtaining image sparse indicates coefficientWherein i=1,2 ...,
3136；According to the different content of the non-smooth block peace sliding block of digital picture, according to being ordered from large to small image sparse table
Show factor alpha0；
(8) judgement, analytic process:
(8.1) according to formula yi=ΦB·xiWithCalculate separately the measurement vector y of i-th of image blockiWith
The measurement data Y of whole image X；According to formulaThe energy of i-th of image block is calculated,
And combine the relationship of its energy size and threshold valueDetermine the data area of i-th of image block: if Ei≤
T then belongs to the smooth block of digital picture；If Ei> T then belongs to the non-smooth block of digital picture；
The calculating process of the measurement data Y of (8.1.1) image X is as follows:
The measurement vector y of i-th of image blockiAre as follows: yi=ΦB·xi, wherein i=1,2 ..., B2；
The then measurement data Y of whole image X are as follows:
(8.1.2) determines that the non-smooth block peace sliding block process of digital picture is as follows:
According to BCS measured value yi, calculate the energy of i-th of image blockWherein i
=1,2 ..., B × B (B is piecemeal size), and according toDivide the image into smooth and non-partes glabra
Point；
(8.2) searching and xiSimilar L image block obtains being indicated with the linear combination of L similar block to present image
Block xiEstimated value xi′:
Wherein,Indicating the design factor of similar block, its value is directly proportional to the noise variance of image in the area,
The noise variance is estimated with steady median operator；
(8.3) Hadamard calculation matrix ΦBSubstitute into formula:
Wherein β is iterative parameter, and value range is | Ri-Ri| < β, 0 < β < 1 update the piecemeal of current survey image
Calculated valueK is iteration optimization number；
(8.4) according to formulaIt is solved with Lasso algorithm
To αiInitial solutionAnd substitute into formulaUpdate image block redundant dictionary
Coefficient, and substitute into following formula:
Seek the sparse bayesian learning coefficient of detection image block
Wherein, ΦBFor piecemeal Hadamard calculation matrix, y is measured value, and d is the atom in dictionary D, and R is iterative calculation
Error, α are dictionary coefficient, i=1,2 ..., n, n=3136；
(8.5) analysis judgement output: ifThen i-th of image block contains secret data in jth block dictionary,
Secret data position is xi；IfThen i-th of image block is free of secret data in jth block dictionary, is clean original
Beginning image block；
(9) iteration is searched for, until detecting complete image data.
The present invention passes through experiment, simulation, the specific implementation process is as follows:
(1) experiment condition and parameter
There are 4000 several original color images in the image library that this experiment uses, wherein the image of the airspace BMP, GIF format
Each 1000 several, jpeg image 2000 several.10 kinds (5 kinds of airspace steganography tools, 5 kinds of DCT domain steganography tools) are listed no with table 1
With obtaining covering each 1000 width of close image after steganography tool embedding information.1000 width of spatial domain picture clean image (its is used when training
Middle each 500 width of BMP, GIF), 2500 width cover close image with 5 kinds of airspace steganography tool insertion secrets are resulting；Similarly, DCT domain
The clean image of 1000 width of image and 2500 width are resulting after being embedded in secret informations with 5 kinds of DCT domain steganography tools to be covered close image and is used as
Training, residual image be used as test image, and with Photoshop made spatial domain picture and DCT domain image it is smoothed denoising, it is sharp
Change, reduce, shearing (only doing spatial domain picture), recompression (only doing JPG image) each 100 width, for test.
The conceptual illustration of some detection performance indexs utilized when experiment is as follows:
Positive detection PD (Positive Detect ion): stego-image is correctly identified.
Feminine gender detection ND (Negat ive Detect io n): original image is correctly identified.
False positive (or false-alarm) FP (False Posit ive): original image is mistaken for stego-image.
False negative (or false dismissal) FN (False Negat iv e): stego-image is mistaken for original image.
1 steganography tool of table and insertion rate
Table 2 lists the steganalysis result realized to different steganographic algorithms.In that, airspace is positive for operation domain in table
Verification and measurement ratio is exactly the average positive verification and measurement ratio of 5 kinds of airspace steganography modes same, and DCT domain positive detection rate is 5 kinds of DCT domain steganography
The average positive detection rate of mode.
The different steganography modes of table 2, different operation domain, different disposal and original image average PD, ND, FN and FP rate
Table 3, which is listed, to be realized under the conditions of different steganographic algorithms containing close and not stego-image judging result.The explanation of table 3,
The secret location of pixels of middle stego-image can specifically be demarcated in detected image block, illustrate steganalysis of the present invention
Precision reached Pixel-level.
To containing close and not stego-image judgement position result under the conditions of the different steganographic algorithms of table 3
Fig. 4 is to realize the experimental result indicated the dictionary of different images by this method in image library.And Fig. 5 is
The experimental result schematic diagram for the Secret Image position that stego-image can determine by this method dictionary learning in image library.
The implementation of the present invention is not limited to this, under the premise of above-mentioned basic fundamental thought of the invention, according to this field
Ordinary technical knowledge and customary means make the modification, replacement or change of other diversified forms to the content of present invention, all fall within
Within rights protection scope of the present invention.
Claims (3)
1. a kind of robust general steganalysis method of picture material perception, it is characterised in that the following steps are included:
(1) the size N, piecemeal size B and measured value size M of digital picture X are set；
(2) 9/7-Haar DWT sparse matrix Ψ is determined:
(3) the Hadamard calculation matrix Φ of BCS is determinedB:
(4) digital picture is divided into B × B block, and according to Hadamard calculation matrix ΦBCalculate the measurement vector y of i-th of image blocki
With the measurement data Y of whole image X:
Wherein, yi=ΦB·xi, i=1,2 ..., B2；
The then measurement data Y of whole image X are as follows:
(5) according to the energy of each image block of BCS measured valueIt calculatesAnd judge the measurement vector y of i-th of image blockiWhether threshold value T=is more than or less than
0.4E, to divide the image into smooth and non-smooth perceptually result；
(6) training process；
(7) assorting process；
(8) judgement, analytic process；
(9) iteration is searched for, until detecting complete image data；
Specific step is as follows for step (6) training process:
(6.1) according to the different digital perception of content of step (5) as a result, determining RDWT dictionary and orthogonal DCT dictionary: forming dictionary
D=[D1, D2 ..., D13]；
2 grades of RDWT dictionary formats in dictionary D are as follows:
Wherein, symbolIt indicates inner product operation, is a kind of operation of element multiplication between matrix；
The form of orthogonal DCT dictionary are as follows:
(6.2) the sparse initial value of dictionary is calculatedAnd calculate image block xiInitial value:
Described image block xiThe specific calculating process of initial value is as follows: setting diFor middle dictionary DiAtom, RiIt is irregular for iteration: Ri=
||ΦBxi-ΦBxi-1||2=| | yi-ΦBxi-1||2,The sparse initial value of dictionaryIt is dictionary atom diWith repeatedly
For irregular RiInner product, the subscript 0 of symbol indicates the 0th value, that is, initial value；
It obtains in this way, image block xiInitial valueForm are as follows:
Wherein, n is data vector length and dictionary size, i=1,2 ..., n；The subscript 0 of symbol indicates the 0th value, that is,
Initial value；
(6.3) image block x is calculatedi=D αi,iAnd according to formula: Ri=| | ΦBxi-ΦBxi-1||2=| | yi-ΦBxi-1||2Iteration
Training updates dictionary D；
Calculating process similar step (6.2), way of realization is as follows:
Specific step is as follows for step (7) assorting process:
(7.1) according to formula yi=ΦB·xi, calculate the measured value y of i-th of image blockiAnd its energyAnd combine the relationship of energy and threshold valueDetermine digital picture
The data area of different content；
(7.2) energy of each image block is calculatedSize: setting the ceiling capacity and least energy of respectively image, meter
The testing time of nomogram pictureEmFor image energy mean value；
(7.3) according to formulaDetermine the measured value y of i-th of image blockiIn jth
The sparse coefficient of a dictionaryWherein j=1,2 ..., 13；
Way of realization it is as follows:
(7.4) it iterates to calculate, obtaining image sparse indicates coefficientWherein i=1,2 ..., 3136；It presses
According to the different content of the non-smooth block peace sliding block of digital picture, coefficient is indicated according to image sparse is ordered from large to small
α0；
Specific step is as follows for the step (8):
(8.1) according to formula yi=ΦB·xiWithCalculate separately the measurement vector y of i-th of image blockiWith it is entire
The measurement data Y of image X；According to formulaThe energy of i-th of image block is calculated, and is tied
Close the relationship of its energy size and threshold valueDetermine the data area of i-th of image block: if Ei≤ T, then
Belong to the smooth block of digital picture；If Ei> T then belongs to the non-smooth block of digital picture；
(8.2) searching and xiSimilar L image block obtains being indicated with the linear combination of L similar block to current image block xi's
Estimated value xi′:
Wherein,Indicate the design factor of similar block, its value is directly proportional to the noise variance of image in the area, described
Noise variance is estimated with steady median operator；
(8.3) Hadamard calculation matrix ΦBSubstitute into formula:
Wherein β is iterative parameter, and value range is | Ri+1-Ri| < β, 0 < β < 1, the piecemeal for updating current survey image calculate
ValueK is iteration optimization number；
(8.4) according to formulaIt solves to obtain α with Lasso algorithmi
Initial solutionAnd substitute into formulaUpdate image block redundant dictionary system
Number, and substitute into following formula:
Seek the sparse bayesian learning coefficient of detection image block
Wherein, ΦBFor piecemeal Hadamard calculation matrix, y is measured value, and d is the atom in dictionary D, and R is the error of iterative calculation,
α is dictionary coefficient, i=1,2 ..., n, n=3136；
(8.5) analysis judgement output: ifThen i-th of image block contains secret data in jth block dictionary, secret
Data Position is xi；IfThen i-th of image block is free of secret data in jth block dictionary, is clean original graph
As block.
2. according to the method described in claim 1, it is characterized by: in the step (8.1) the measurement data Y of image X meter
Calculation process is as follows:
The measurement vector y of i-th of image blockiAre as follows: yi=ΦB·xi, wherein i=1,2 ..., B2；
The then measurement data Y of whole image X are as follows:
3. according to the method described in claim 1, it is characterized by: determining the non-smooth block of digital picture in the step (8.1)
It is as follows with smooth block process:
According to BCS measured value yi, calculate the energy of i-th of image blockWherein i=1,
2 ..., B × B, and according toDivide the image into smooth and non-smooth.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610567405.2A CN106228505B (en) | 2016-07-15 | 2016-07-15 | A kind of robust general steganalysis method of picture material perception |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610567405.2A CN106228505B (en) | 2016-07-15 | 2016-07-15 | A kind of robust general steganalysis method of picture material perception |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106228505A CN106228505A (en) | 2016-12-14 |
| CN106228505B true CN106228505B (en) | 2019-06-18 |
Family
ID=57531833
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610567405.2A Expired - Fee Related CN106228505B (en) | 2016-07-15 | 2016-07-15 | A kind of robust general steganalysis method of picture material perception |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106228505B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106909819A (en) * | 2016-12-26 | 2017-06-30 | 浙江海洋大学 | It is a kind of to be suitable to the image information steganography method of intelligent terminal |
| CN106991652A (en) * | 2017-03-03 | 2017-07-28 | 杭州电子科技大学 | Degree of rarefication constrains the coloured image restorative procedure with dictionary atom size adaptation |
| CN108307086B (en) * | 2018-01-26 | 2019-05-07 | 广东技术师范学院 | A kind of method and device of image latent writing |
| CN108346125B (en) * | 2018-03-15 | 2021-10-08 | 中山大学 | Airspace image steganography method and system based on generation countermeasure network |
| CN109214973B (en) * | 2018-08-24 | 2020-10-27 | 中国科学技术大学 | Method for generating countermeasure security carrier aiming at steganalysis neural network |
| CN109168005A (en) * | 2018-09-18 | 2019-01-08 | 天津大学 | A kind of two-dimensional discrete wavelet conversion construction design method improving storage efficiency |
| CN111768326B (en) * | 2020-04-03 | 2023-08-25 | 南京信息工程大学 | High-capacity data protection method based on GAN (gas-insulated gate bipolar transistor) amplified image foreground object |
| CN111932646B (en) * | 2020-07-16 | 2022-06-21 | 电子科技大学 | Image processing method for resisting attack |
| CN113469869B (en) * | 2021-09-03 | 2021-11-12 | 武汉华工安鼎信息技术有限责任公司 | Image management method and device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102546126A (en) * | 2012-01-05 | 2012-07-04 | 南京师范大学 | Information hiding and extraction method based on multi-parameter change normal similar-orthogonal matrix |
| CN102917227A (en) * | 2012-10-29 | 2013-02-06 | 山东省计算中心 | Compressive sensing-based adaptive video information hiding method |
| CN103152608A (en) * | 2013-02-28 | 2013-06-12 | 广东技术师范学院 | Method for realizing digital fingerprint encrypted video multicast transmission |
| CN103247028A (en) * | 2013-03-19 | 2013-08-14 | 广东技术师范学院 | Multi-hypothesis prediction block compressed sensing image processing method |
-
2016
- 2016-07-15 CN CN201610567405.2A patent/CN106228505B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102546126A (en) * | 2012-01-05 | 2012-07-04 | 南京师范大学 | Information hiding and extraction method based on multi-parameter change normal similar-orthogonal matrix |
| CN102917227A (en) * | 2012-10-29 | 2013-02-06 | 山东省计算中心 | Compressive sensing-based adaptive video information hiding method |
| CN103152608A (en) * | 2013-02-28 | 2013-06-12 | 广东技术师范学院 | Method for realizing digital fingerprint encrypted video multicast transmission |
| CN103247028A (en) * | 2013-03-19 | 2013-08-14 | 广东技术师范学院 | Multi-hypothesis prediction block compressed sensing image processing method |
Non-Patent Citations (5)
| Title |
|---|
| Cognitive Computation of Compressed Sensing for Watermark Signal Measurement;Huimin Zhao;《Cognitive Computation》;20150924;全文 |
| On the Security and Robustness with Fingerprint Watermarking Signal via Compressed Sensing;Huimin Zhao;《International Journal of Security and Its Applications》;20150131;全文 |
| 一种分块压缩感知变采样率的指纹图像水印算法;赵慧民;《中山大学学报(自然科学版)》;20140930;全文 |
| 压缩感知的自适应冗余字典及其图像恢复方法研究;蒋业文;《中山大学学报( 自然科学版)》;20121130;全文 |
| 结合字典稀疏表示和非局部相似性的自适应压缩成像算法;练秋生;《电子学报》;20120731;全文 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106228505A (en) | 2016-12-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106228505B (en) | A kind of robust general steganalysis method of picture material perception | |
| Qureshi et al. | A bibliography of pixel-based blind image forgery detection techniques | |
| CN110443293B (en) | Zero sample image classification method for generating confrontation network text reconstruction based on double discrimination | |
| Shojae Chaeikar et al. | PSW statistical LSB image steganalysis | |
| Cui et al. | An early fire detection method based on smoke texture analysis and discrimination | |
| CN106530200A (en) | Deep-learning-model-based steganography image detection method and system | |
| AU2017201281B2 (en) | Identifying matching images | |
| Sun et al. | Secure and robust image hashing via compressive sensing | |
| Zong et al. | Blind image steganalysis based on wavelet coefficient correlation | |
| CN104661037B (en) | The detection method and system that compression image quantization table is distorted | |
| Zhao et al. | Tampered region detection of inpainting JPEG images | |
| Sasirekha et al. | Enhanced techniques for PDF image segmentation and text extraction | |
| Wang et al. | A statistical feature based approach to distinguish PRCG from photographs | |
| CN110457996A (en) | Moving Objects in Video Sequences based on VGG-11 convolutional neural networks distorts evidence collecting method | |
| Shenbagavalli et al. | Aiding the digital mammogram for detecting the breast cancer using shearlet transform and neural network | |
| Xu et al. | Detecting image splicing using merged features in chroma space | |
| CN106530199B (en) | Multimedia integration steganalysis method based on window type hypothesis testing | |
| CN105512677B (en) | Classifying Method in Remote Sensing Image based on Hash coding | |
| CN114694090A (en) | Campus abnormal behavior detection method based on improved PBAS algorithm and YOLOv5 | |
| CN113887652A (en) | Remote sensing image dim target detection method based on form and multi-example learning | |
| Li et al. | An image steganalysis method based on characteristic function moments and PCA | |
| Yin et al. | Hyperspectral image classification using wavelet packet analysis and gray prediction model | |
| Bartels et al. | Wavelet packets and co-occurrence matrices for texture-based image segmentation | |
| Zhu et al. | Recaptured image detection through multi-resolution residual-based correlation coefficients | |
| Abdalmajeed et al. | Using the natural scenes’ edges for assessing image quality blindly and efficiently |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CP03 | Change of name, title or address |
Address after: 510665 No. 293 Shipai Zhongshan Avenue, Tianhe District, Guangzhou City, Guangdong Province Patentee after: Guangdong Normal University of Technology Address before: 510665 No. 293 Zhongshan Avenue, Guangzhou, Guangdong Province Patentee before: Guangdong Technical Normal College |
|
| CP03 | Change of name, title or address | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190618 Termination date: 20210715 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |