The content of the invention
It is an object of the invention to provide a kind of LSB based on gray level co-occurrence matrixes statistical nature replaces steganalysis method,
The dimension of characteristic vector is the method reduced, " dimension disaster " present in prior art is solved and accuracy of detection is low asks
Topic.
The technical solution adopted in the present invention is that a kind of LSB based on gray level co-occurrence matrixes statistical nature replaces steganography point
Analysis method, implements according to following steps:
Step 1, decomposition image bit-plane
For the image I that gray level is 0-255, the decomposition formula of its image bit-plane is as follows:
Wherein, I (x, y) represent image I in (x, y) place pixel value, BiRepresent i-th bit plane, Bi(x, y) is represented
Bit plane BiIn (x, y) place pixel value, i=1,2 ..., 8, s, t=256,
Step 2, calculating gray level co-occurrence matrixes
2.1) calculate difference matrix Dk:
Wherein, s, t=256, k=1,2 ..., 7;
2.2) calculate difference matrix and matrix:
Wherein, s, t=256, d11,...,dst∈{0,1,2,3,4,5,6,
7};
2.3) calculate MDCo-occurrence matrix
Mathematically, the co-occurrence matrix being defined on the matrix I of a n × m, by arranging a group of side-play amount (Δ x, Δ y)
It is parameterized as follows:
Wherein, the formula of and if only if for δ=1 bracket behind is set up,
M is calculated belowDCo-occurrence matrix G, obtain the matrix of 8 × 8:
Step 3, extraction feature
According to the dependency between image neighbor and the distribution character with matrix, in G, the distribution pattern of element is to become
In integrated distribution, the value of zone line is bigger than the value of marginal area, then the value of zone line is regarded as the main letter of image
Breath;
Step 4, classification
If secret information is embedded into the minimum bit plane of image, the dependency between bit plane will change, this
Change will be shown on co-occurrence matrix G, while can prove whether hide secret information in image as evidence.
The invention has the beneficial effects as follows:
1) method of the present invention make use of the dependency between image bit-plane, from being total to for the difference matrix of different bit planes
Feature is extracted in raw matrix, the dimension of characteristic vector is reduced, be efficiently avoid " dimension disaster ", and reached expected effect
Really.
2) algorithm has higher verification and measurement ratio, and stability preferably, is robust to JPEG compression, addition noise, filtering,
With preferable generalization ability, computation complexity is low.
3) method of the present invention can be used to detect that LSB replaces steganography that the effect spread that has to protecting internet information is tieed up
Shield the Internet public order has effective meaning.
Specific embodiment
The present invention is described in detail with reference to the accompanying drawings and detailed description.
LSB of the present invention based on gray level co-occurrence matrixes statistical nature replaces steganalysis method, referring to Fig. 1, including image
Bit Plane Decomposition, calculating gray level co-occurrence matrixes, the selection of feature and extraction, classifying step, are described in detail below:
Step 1, decomposition image bit-plane
For the image I that gray level is 0-255, the decomposition formula of its image bit-plane is as follows:
Wherein, I (x, y) represent image I in (x, y) place pixel value, BiRepresent i-th bit plane, Bi(x, y) is represented
Bit plane BiIn (x, y) place pixel value, i=1,2 ..., 8, s, t=256,
Step 2, calculating gray level co-occurrence matrixes
2.1) calculate difference matrix Dk:
Wherein, s, t=256, k=1,2 ..., 7;
2.2) calculate difference matrix and matrix:
Wherein, s, t=256, d11,...,dst∈{0,1,2,3,4,5,6,
7};
2.3) calculate MDCo-occurrence matrix
Mathematically, the co-occurrence matrix being defined on the matrix I of a n × m, by arranging a group of side-play amount (Δ x, Δ y)
It is parameterized as follows:
Wherein, the formula of and if only if for δ=1 bracket behind is set up.
In the method for the invention, the parameter of co-occurrence matrix selects as follows:The offset parameter Δ x of co-occurrence matrix is set to 1,
Δ y is set to 0, that is to say, that only consider the adjacent element of horizontal direction, the co-occurrence matrix for being generated when co-occurrence matrix is calculated
It is unsymmetrical matrix,
M is calculated belowDCo-occurrence matrix G, obtain the matrix of 8 × 8:
Step 3, extraction feature
According to the dependency between image neighbor and the distribution character with matrix, in G, the distribution pattern of element is to become
In integrated distribution, that is to say, that the value of zone line is bigger than the value of marginal area, then regard the value of zone line as image
Main information, before and after steganography image essence occur minor variations, the main information of image also can be changed therewith, with this
As statistical significance feature.That is, after secret information is embedded into minimum bit plane, minimum bit plane and remaining seven
Dependency between bit plane changes, and this change is described by the value of the zone line of G, therefore, select the centre of G
16 elements in region are used as characteristic vector:
F={ g33,g34,g35,g36,g43,g44,g45,g46,g53,g54,g55,g56,g63,g64,g65,g66,
In order to illustrate that characteristic vector f is sensitive to steganography, tested in experiment respectively 1600 secondary carrier images and
1600 secondary hidden close images, Fig. 2 and Fig. 3 show the distribution of the assembly average of the gray level co-occurrence matrixes G for generating in aforementioned manners
Pattern.
Fig. 2 represents dividing for the assembly average of the gray level co-occurrence matrixes G for being generated to 1600 secondary carrier images in aforementioned manners
Cloth pattern, Fig. 3 represent the distribution of the assembly average of the gray level co-occurrence matrixes G for being generated to 1600 secondary hidden close images in aforementioned manners
Pattern.As can be seen that the assembly average of carrier image and the gray level co-occurrence matrixes G corresponding to hidden close image from two figures
Distribution pattern all shows as the significant characteristic of zone line, and before and after data are embedded, the value of zone line element can occur
Change, and the amount of this change is different, this different patterns of change can be used for distinguishing whether have steganography information
Exist.
Step 4, classification
If secret information is embedded into the minimum bit plane of image, the dependency between bit plane will change, this
Change will be shown on co-occurrence matrix G, while can prove whether hide secret information in image as evidence.
Characteristic vector f is defined as characteristic of division, using LS-SVM as grader distinguishing carrier image and hidden close
Image, here select RBF RBF as kernel function.
The simulation experiment result of the inventive method
Image used in experiment is from Colombia's image library [12] and UCID image libraries [13], the image of these tests
Different shading values, texture and details are contained, and the hidden close image of different embedded rates are generated to assess this using these images
Bright method, the foundation of sample set are as follows:
First, 1600 test images (size is 256 × 256) are converted to the gray level image of bitmap format, generate and carry
Body image pattern collection, is defined as IC;Secondly, the minimum bit plane in carrier image is embedded in secret information and generates hidden close image, embedding
The size for entering rate is 12%, 25% and 50% respectively, obtains three hidden Mi Tuxiangyangbenji, is defined as IS-1, IS-2And IS-3。
1) test of accuracy of detection
A) false positive rate and false negative rate
For the accuracy of detection of quantitative analyses the inventive method, false positive rate and false negative rate are defined as follows:
Here, | | | | represent the operation of the radix of set of computations.
In order to discuss the performance of the inventive method, 30 false negative rates and false positive rate are tested respectively.
In experiment, from ICIn randomly select 400 carrier images, from IS-3In randomly select 400 hidden close images as instruction
Practice sample, from IS-3Randomly select 10n hidden close image to be surveyed as test specimens originally in-{ the 400 hidden close images chosen }
Examination PFN, from ICRandomly select 10n carrier image to be tested as test specimens originally in-{ 400 carrier images chosen }
PFP, n=1 ... here, 30.Experimental result shows in figures 4 and 5.
Fig. 4 and Fig. 5 show and compare the distribution situation of false negative rate and false positive rate respectively, it can be seen that
The false negative rate and false positive rate of the inventive method is less than 5%.
From experimental result as can be seen that the testing result of document [7] presents extreme phenomenon, i.e., when training sample is different
When, testing result or very high value 99.5% is reached, or be 0, present unstability.In false positive rate and false negative rate
In experiment, in order to avoid the extreme phenomenon presented as the verification and measurement ratio that the unstability of algorithm in document [7] causes, 30 experiment institutes
The training sample for using is identical, and different from other experiments, the training sample for using is random selection to this point every time
's.
B) comparative experimentss
In order to illustrate that the inventive method has superiority in terms of accuracy of detection than the existing method based on co-occurrence matrix,
Experiment is compared, and is tested and true negative rate P has been counted in the case of different embedded rates 12%, 25% and 50%TN=1-PFP
With True Positive Rate PTP=1-PFN, as a result list in table 1.In table 1, PT=(PTN+PTP)/2。
From ICIn randomly select 400 carrier images, from IS-1In randomly select 400 hidden close images as training sample,
Then from IC200 carrier images are randomly selected as test sample in-{ 400 carrier images chosen }, using LS-
SVM is classified, and obtains false positive rate PFP.Repeat this process 10 times, the meansigma methodss of statistical experiment result are used as final
PFP, it is shown that its corresponding true negative rate P in table 1TN=1-PFP.For the stego image that embedded rate is 25% and 50%, survey
Examination process is identical with mentioned above, and all of test result is all displayed in table 1.
From ICIn randomly select 400 carrier images, from IS-1In randomly select 400 hidden close images as training sample,
Then from IS-1200 hidden close images are randomly selected in-{ the 400 hidden close images chosen } as test sample, using same
Test PFPThe same method obtains false negative rate PFN, its corresponding True Positive Rate P displayed in Table 1TP=1-PFN。
The comparative result of table 1, accuracy of detection
From table 1 it follows that compare with document [7] with document [9], the true negative rate of the inventive method and True Positive Rate
It is gratifying.
2) stability analyses
In order to whether the algorithm for testing the present invention is applied to different sample sets, that is to say, that whether random selection sample set
Impact to experimental result is negligible, and defines " stability " this concept.Here stability is referred to for difference
Sample set, the verification and measurement ratio of algorithm changes presented stability in the range of very little, and preferable steganalysis algorithm should
Should all be stable for different sample sets.In order to prove the stability of inventive algorithm, respectively different embedded rates 12%,
10 experiment tests are carried out under 25% and 50%, wherein the training sample and test sample of experiment is all random selection every time
, detailed results are as shown in table 2.
The comparison of table 2, stability experiment testing result
From Table 2, it can be seen that the inventive method is more stable than document [7] and document [9], the excursion of verification and measurement ratio is less than
10%, rather than as for example high verification and measurement ratio or 0 of document [7] extreme case that occurs, while the verification and measurement ratio excursion of document [9]
Less than 27%.Result in table shows that the feature that the inventive method is extracted is stable to classification.
3) robust analysis
This experiment is robust for testing the inventive method to operating the caused distortion that is not intended to by content retentivity,
Wherein content retentivity operation includes JPEG compression, adds noise, filtering etc..
The embedded rate of hidden close image that experiment generation training sample is used is 50%.Test sample is respectively from ICIn it is random
Choose 200 carrier images and obtained with filtering by adding noise, JPEG compression.In order to quantitatively analyze inventive algorithm
Robustness, defines the concept of " accuracy ", accuracy refer to algorithm will live through content keep the image detection of operation into
The probability of non-hidden close image, its formula are as follows:
Here, | | | | represent the operation of the radix of set of computations.
It is shown that the meansigma methodss of 10 experiment accuracy results, the training sample tested every time and test sample in table 3
All it is randomly selected.
Table 3, the accuracy result to the operation of content retentivity
From table 3 it is observed that operating to content retentivity, the method for the present invention presents gratifying detection knot
Really, this means that the inventive method is robust to the operation of content retentivity.
4) generalization ability analysis
In general, different image libraries has a great impact to the performance of steganalysis algorithm, in real world, comes
It is very big from not homologous image difference, therefore a kind of grader can not possibly be trained suitable for all of image library on the Internet
[11].The generalization ability of steganalysis algorithm refers to the usability to different images data base, and high generalization ability is referred to works as
Image training sample and test sample from different sources when, the verification and measurement ratio of algorithm changes in the range of very little.
In order to test the generalization ability of the inventive method, in experiment image from UCID and Colombia's image library, 1000
Open UCID images (size is 256 × 256) and 1000 Colombia's images (size is 256 × 256) are converted to bitmap lattice
The gray level image of formula, therefrom respectively chooses the embedded secret information of half, and the rate that is embedded in is 50%.The experiment contains four groups of son experiments,
Per group of experiment repeats 10 times, randomly chooses sample image every time, and the ratio of wherein training sample and test sample is 4:
1。
Table 4, for the generalization ability analysis result in different images storehouse
Experimentation is as follows:
Training sample and test sample both are from UCID image libraries (UCID → UCID)
, from UCID image libraries, test sample is from Colombia's image library (UCID → Columbia) for training sample
Training sample and test sample both are from Colombia's image library (Columbia → Columbia)
, from Colombia's image library, test sample is from UCID image libraries (Columbia → UCID) for training sample
Testing result is as shown in table 4:
As can be seen from Table 4, when training sample and test sample are from different sources, models of the PTN and PTP in very little
Enclose interior change, it means that the inventive method has gratifying generalization ability.
5) computation complexity analysis
Computation complexity refers to extract the time that feature is consumed from training sample and test sample image.In experiment,
Training sample is made up of 400 carrier images and 400 hidden close images, and test sample is made up of 200 carrier images
's.
Table 5, computation complexity compares
|
The method of the present invention |
Method in document [7] |
Method in document [9] |
Generate the time (second) that training sample is consumed |
22.191 |
37.138 |
38.754 |
Generate the time (second) that test sample is consumed |
5.577 |
8.867 |
9.637 |
Experimental result is as shown in table 5, it can be seen that the inventive method is presented relative to document [7] and document [9]
High efficiency.
List of references
[1] Lignum Sappan seven. Steganography:Encrypt in figure .http sprouting://www.guokr.com/article/3741/
.2010.
[2] Johnson N F, Jajodia S.Steganalysis of images created using current
steganography software[J].In Proceedings of the 2th International Workshop on
Information Hiding, Portland, Oregon, USA, April 15-17,1998,273-289.
[3] button heart Xin. Information hiding and digital watermarking [M]. Beijing:Publishing house of Beijing University of Post & Telecommunication, 2004.
[4] Fridrich J, Goljan M.On Estimation of Secret Message Length in LSB
Steganography in Spatial Domain[J].In IS&T/SPIE Electronic Imaging:Security,
Steganography, and Watermarking of Multimedia Contents VI, San Jose, California,
USA, 19-22January, 2004,23-34.
[5]Ker AD.A general framework for the structural steganalysis of LSB
replacement[J].In Proceedings of the 7thInformation Hiding Workshop,
Barcelona, 6-8June, LNCS3727,2005,296-311.
[6]Ker A D.Fourth-order Structural Steganalysis and Analysis of Cover
Assumptions[J].In IS&T/SPIE Electronic Imaging:Security, Steganography, and
Watermarking of Multimedia Contents VIII, San Jose, USA, 23January, 2006,1-14.
[7] Abolghasemi M, Aghainia H, Faez K, Mehrabi M A.LSB Data Hiding
Detection Based on Gray Level Co-Occurrence Matrix(GLCM)[J].International
Symposium on Telecommunications, 2008,656-659.
[8] Kekre H B, Athawale AA, Patki S A.Steganalysis of LSB Embedded
Images Using Gray Level Co-Occurrence Matrix[J].International Journal of
Image Processing, 2011,5 (1):36-45.
[9] Deng Qianlan, Lin Jiajun. the image latent writing parser [J] based on co-occurrence matrix. microcomputer information (observing and controlling
Automatization), 2009,25 (12):6-8.
[10] Ghanbari S, Keshtegary M, Ghanbari N.New Steganalysis Method using
GLCM and Neural Network [J] .International Journal of Computer Applications,
2012,42 (7):46-52.
[11] NG W W Y, He Z M, Chan P P K, Yeung D S.Blind Steganalysis with High
Generalization Capability for Different Image Databases Using L-GEM[J]
.Proceedings of the 2011International Conference on Machine Learning and
Cybernetics, Guilin, 10-13July, 2011,1690-1695.
[12] T.T.Ng, S.F.Chang, J.Hsu, M.Pepeljugoski, " Columbia Photographic
Images and Photorealistic Computer Graphics Dataset ", Department of Electrical
Engineering, Columbia University, Advent Technical Report #205-2004-5, Feb.2005.
[13] UCID, http://www-staff.lboro.ac.uk/~cogs/datasets/UCID/ucid.html.