JP2009524078A - How to identify marked content - Google Patents

Abstract

A method for identifying marked content is provided.
Briefly, according to one embodiment, a method for identifying marked content is described.
[Selection] Figure 1

Description

The present application relates to the classification or identification of content, such as marked content.

Digital data hiding technology is a field that has been actively studied in recent years, and various data hiding methods have been proposed. Some methods aim at content protection and / or authentication, while others aim at secret communication. Data hiding classified as the latter is referred to herein as steganography.

The subject matter of the present invention is set forth with particularity in the appended claims and is claimed explicitly. However, the claimed subject matter will be best understood by reference to the following detailed description and the accompanying drawings, as well as its purpose, features and / or advantages, both in terms of the mechanism and method of operation.

DETAILED DESCRIPTION In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of claimed subject matter. However, one of ordinary skill in the art appreciates that the subject matter recited in the claims can be practiced without the specific details. In other instances, well known methods, procedures, components, and / or circuits have not been described in detail so that the claimed subject matter is not obscured.

In the detailed description that follows, an algorithm and / or symbolic representation of the manipulation of data bits and / or binary digital signals stored in a computing system (such as a computer and / or computing system memory) is described. There is a part. These algorithmic descriptions and / or representations are techniques used by those skilled in the data processing arts to convey their work to others skilled in the art. An algorithm is herein and generally considered a self-consistent sequence of operations and / or similar processes that yields the desired result. The above operations and / or processes may involve physical manipulation of physical quantities. These physical quantities are typically, but not necessarily, in the form of electrical and / or magnetic signals that can be stored, transferred, combined, compared, and / or otherwise manipulated. It may be convenient to refer to these signals as bits, data, values, elements, symbols, characters, terms, numbers, and / or numbers, primarily in terms of general usage. It should be understood, however, that these and similar terms are all associated with the appropriate physical quantities and are merely convenient labels. Unless stated otherwise, the description using terms such as “processing”, “arithmetic”, “calculation”, and / or “determining” is used herein to be a computing platform, as will be apparent from the following description. Process data represented as physical electronic quantities and / or magnetic quantities and / or other physical quantities in a processor, memory, register, and / or other information storage device, information transmission device, and / or information display device It is understood that it refers to operations and / or processes performed by a computing platform (such as a computer or similar electronic computing device) that transforms.

Digital data hiding technology is a field that has been actively studied in recent years, and various data hiding methods have been proposed. Some methods aim at content protection and / or authentication, while others aim at secret communication. Data hiding classified as the latter is referred to herein as steganography.

Literature: J.M. Fridrich, M.M. Goljan and D.C. Hogea, “Steganalysis of JPEG Images: Breaking the F5 algorithm”, 5th Information Hiding Workshop, 2002, pp. 310-323 (hereinafter “Fridrich et al.”), Fridrich et al. Shows that when the F5 embedding method is applied to generate a stego image, the number of zeros in the block DCT domain of the stego image increases. By using this feature, for example, it can be determined whether or not a hidden message is embedded in the content by the F5 method. There are other inventions relating to steganalysis of a method for concealing specific target data. See, for example, the following literature: Fridrich, M.M. Goljan and R.M. Du, “Detecting LSB steganography in color and gray-scale images”, Magazine of IEEE Multimedia Special Issue on Security, Oct. -Nov. 2001, pp. 22-28, and R.A. Chandamouri and N.C. Memon “Analysis of LSB based image steganography techniques”, Proc. of ICIP 2001, Oct. 7-10, 2001.

Literature: S.M. Lyu and H.M. Farid, “Detecting Hidden Messages Using Higher-Order Statistics and Support-and-Land, and Supplied-Further Machines” It proposes a more general steganalysis method based at least in part on higher-order statistics of images derived from image decomposition by filters. Higher order statistics of wavelet high frequency subbands are extracted as steganalysis features in this approach. Similarly, according to this method, a stego image can be distinguished from a cover image with a certain success rate. The data hiding method handled by this particular steganalysis mainly comprises a least significant bit plane (LSB) modified type steganographic tool.

Literature: K.K. Sullivan, U .; Madhow, S .; Chanderaskaran, and B.C. S. Manjunath, “Steganalysis of Spread Spectrum Data Hiding Exploring Cover Memory”, SPIE 2005, vol. 5681, pp 38-46 (hereinafter “Sullivan et al.”) Proposes a steganalysis method based at least in part on a hidden Markov model. The empirical transition matrix of the test image is formed by such a method. However, the size of the empirical transition matrix is large. For example, in a grayscale image with a bit depth of 8, there are 65536 elements in one image. Therefore, this matrix is not used directly as a feature. The author features that some of the maximum probabilities along the main diagonal of the matrix are selected along with their neighbors, and other probabilities along the main diagonal are selected randomly. Unfortunately, some useful information may be ignored due at least in part to the randomness of the characterization. The data hiding method by Sullivan et al. Mainly relates to the spread spectrum (SS) data hiding method. These latter methods generally do not send as many information bits as the LSB method, but the SS method may be used in connection with, for example, secret communications. In addition, the SS method is known to be more robust than the LSB method. Therefore, it is desirable to study the SS method for steganalysis.

One embodiment of a steganalysis system based at least in part on a Markov chain of thresholded prediction error sets for content such as images, for example, is described below, but the claimed subject matter is scoped in this respect. Is not limited. In this specific embodiment, content samples such as image pixels are predicted by neighboring pixels. For example, a prediction error image is generated by subtracting a prediction value from a pixel value and performing threshold processing. Empirical transition matrices along the horizontal, vertical, and diagonal directions of the Markov chain can be a feature of steganalysis in such embodiments. Analysis of variance techniques such as support vector machines (SVM) or genetic processing may be applied to classification or identification, but in this respect as well, the claimed subject matter is not limited in scope.

Further, in this particular embodiment, the claimed subject matter is not limited in scope to only one embodiment, but is steganarially based at least in part on a Markov chain model of a thresholded prediction error image. A cis system may be applied. Image pixels are predicted by neighboring pixels. The prediction error in this particular embodiment is obtained by subtracting the prediction value from the pixel value. Although the range of the difference value increases, many of the difference values may be concentrated in a relatively narrow range near zero due to the correlation between adjacent pixels of the unmarked image. In this specification, the term marked content refers to content that does not appear to contain such hidden information because the data is hidden. Similarly, unmarked or cover content is content whose data is not concealed. However, large values in the prediction error image may be due at least in part to the image content rather than data concealment. Therefore, a large value in the prediction error image may be reduced or removed by the threshold applied to the prediction error, thereby limiting the dynamic range of the prediction error image.

In this particular embodiment, the prediction error image may be modeled using Markov chains, but the claimed subject matter is not limited in scope in this respect. An empirical transition matrix is calculated and becomes a feature of steganalysis. Probabilities in the matrix may be included in the feature vector because the size of the empirical transition matrix is reduced to a size that is easy to process for the classifier, due at least in part to threshold processing. For feature classification, analysis of variance or other statistical techniques may be applied. For example, SVM processing may be applied with linear and non-linear kernels used for classification, details of which will be described later. Here, the term “ANOVA process” refers to the following process: At least a part of such process by fully distinguishing differences due to statistical fluctuations from differences due to non-statistical fluctuations. A process in which data correlation, segmentation, analysis, classification, or other characterization is performed based on the target.

Although the term steganalysis can have a variety of meanings in this particular embodiment, this term refers to a two-class pattern classification approach. For example, the test image is classified as either a cover image (ie, one that does not hide information) or a stego image or a marked image (one that contains hidden data or a hidden message). Typically, in this particular approach or embodiment, the classification is two-part, but the claimed subject matter is not limited in scope to only using two classifications. Other approaches are possible and are included within the scope of the claimed subject matter. Here, each of the two parts is referred to as feature extraction and pattern classification. In many instances, it is desirable to use the image itself for features in this process, due at least in part to the large amount of information that the image contains. Similarly, however, from a feasibility standpoint, the dimensionality of features may be too high for many classifiers. Therefore, feature extraction may be applied.

For computer vision types, it may be desirable for features to represent the shape and color of an object. In contrast, in this particular embodiment, other characteristics may provide useful information. For example, in steganalysis, it is desirable to have a feature that includes information about changes caused by data hiding with respect to information about the contents of an image.

In general, unmarked images, for example, tend to exhibit specific characteristics such as, for example, being continuous, smooth, and correlated between adjacent pixels. Also, hidden data may be unrelated to the content itself. For example, watermarking may change the continuity for unmarked content because it may introduce some random variation, for example. As a result, the correlation between adjacent pixels, bit planes and image blocks can be reduced. In this particular embodiment, it is desirable to amplify these variations that may occur due to data hiding. This can be accomplished by any of a number of possible approaches, but the claimed subject matter is not limited in scope to a particular approach. However, a specific embodiment for realizing this will be described below.

In this particular embodiment, neighboring pixels may be used to predict the current pixel. In the present embodiment, the prediction may be performed in three directions. Also, in this embodiment, these directions include horizontal, vertical and diagonal directions, but in other embodiments, other directions are possible. For prediction, the prediction error can be estimated or determined by subtracting the predicted pixel value from the original pixel value as shown in (1).

e _h (i, j) = x (i + 1, j) −x (i, j)
e _v (i, j) = x (i, j + 1) −x (i, j) (1)
e _d (i, j) = x (i + 1, j + 1) −x (i, j)

In the equation, e _h (i, j) is the prediction error of the horizontal pixel (i, j), e _v (i, j) is the prediction error of the vertical pixel (i, j), and _ed ( i, j) indicate prediction errors of the pixel (i, j) in the diagonal direction. In the present embodiment, three prediction errors are estimated in this way for the pixels of the image. Thus, in the prediction error, three prediction error images respectively indicated by E _h , E _v and E _d in this specification are formed.

Distortion caused by data hiding is usually observed to be small compared to pixel differences associated with various objects in the image, for example. Otherwise, when inspected by the human eye, the distortion itself suggests hidden data, which can compromise confidential communications. Thus, large prediction errors tend to reflect more on image content rather than hidden data. In this particular embodiment, a threshold こ の is employed to address this, and the prediction error is adjusted according to the following rules:

Of course, the scope of the claimed subject matter is not limited to this particular approach. Many other techniques for dealing with prediction errors are possible and are within the scope of the claimed subject matter. Similarly, for example, the wrinkles according to a particular embodiment may not have a fixed value, and may vary, for example, with time, location, and a number of other potential factors.

Nevertheless, in this embodiment, a large prediction error may be zero. In other words, image pixels or other content samples may be considered smooth from a data hiding perspective. Further, in this specific example, the range of the prediction error image is [−T, T], and can take 2 ^* T + 1 values.

Similarly, in the present embodiment, for example, a two-dimensional Markov chain model rather than a one-dimensional model is applied to a prediction error image subjected to threshold processing. FIG. 1A is a schematic diagram illustrating an embodiment of a transition model of a horizontal prediction error image E _h in which a Markov chain is modeled, for example, in the horizontal direction. 1B and 1C are schematic diagrams illustrating corresponding embodiments of E _v and E _d , respectively. As suggested above and as described in more detail below, the elements of the empirical transition matrix of E _h , E _v and E _d of this embodiment are used as features. 1A to 1C, one circle represents one pixel. These figures show an 8 × 8 size image. The arrow represents a change in the state of the Markov chain.

Various techniques can be used to analyze the data in various situations. Here, the term “ANOVA” is used to refer to the following processes or technologies: As a result of applying the above processes or technologies, differences resulting from statistical fluctuations are differences resulting from non-statistical fluctuations. A process or technique that is sufficiently distinguished from and that correlates, segments, classifies, analyzes, or otherwise characterizes data based at least in part on the application of such process or technique. Examples include artificial intelligence technology and processing, neutral networks, genetic processing, heuristics, support vector machines (SVM), etc., but these do not limit the scope of the claimed subject matter. .

The claimed subject matter is not limited to SVM or SVM processing, but may be a convenient technique for two-class classification. For example, see the following literature: C.I. Cortes and V.M. Vapnik, “Support-vector networks,” in Machine Learning, 20, 273-297, Kluwer Academic Publishers, 1995. For example, SVM can be used to handle linear and non-linear cases or situations. When linear separation is possible, for example, an SVM classifier is applied to obtain a hyperplane that separates positive patterns from negative patterns. As an example, training data pairs {y _i , ω _l }, i = 1,. . . , L (where y _i is a feature vector and ω _l = ± 1 for a positive / negative pattern).

In this particular embodiment, the linear support vector processing can be formulated as follows: If a separation hyperplane exists, the training data satisfies the following constraints.

Similarly, the Lagrangian equation is constructed as follows.

Where α _i is a positive Lagrangian multiplier introduced into the inequality constraints (here (3) and (4)). For w and b, the gradient L gives:

In this embodiment, the sample z from the test data may be classified using w and b by training the SVM classifier. Example, in one embodiment, when the w t ^{z +} b above zero, may be classified as having a message hidden images. Alternatively, it may be classified as not including a hidden message. Of course, this is a specific embodiment, and the claimed subject matter is not limited in scope in this respect. For example, the definition regarding positive, negative, or functional form may be changed according to various factors and states.

When non-linearly separable, the “learning machine” maps the input feature vector to a higher dimensional space where the linear hyperplane is potentially located. In this embodiment, the conversion from the non-linear feature space to the linear high-dimensional space may be performed using a kernel function. Examples of kernels include linear, polynomial, radial basis functions, and sigmoids. In this particular embodiment, for example, a linear kernel may be used for linear SVM processing. Similarly, other kernels may be used for nonlinear SVM processing.

Although a specific system for identifying or classifying marked content, such as images, has been specified, it is desirable to build and evaluate performance, but this is still only one specific embodiment for illustration. Rather, it is noted that the claimed subject matter is not limited in scope to this particular embodiment or approach.

For the evaluation, Vision Research Lab, University of California , the 2812 pieces of the image from the web site of Santa Barbara (see http://vision.ece.ucsb.edu/~sullivak/Research_imgs/), further, CorelDRAW Version Downloaded 1096 sample images contained in 10.0 software CD # 3 (see www.core.com ).

Thus, 3908 images were used as a test image data set. The color image was converted to a grayscale image by applying an irreversible color transformation as shown in (7) below (see for example M. Rabani and R. Joshi, “An Overview of”). the JPEG2000 Still Image Compression Standard ", Signal Processing: Image Communication 17 (2002) 3-48).

Y = 0.299R + 0.587G + 0.114B (7)

The following normal data hiding methods were applied to these images: see, for example, Cox et al. Non-blind SS data concealment method (see IJ Cox, J Kilian, T. Leighton and T. Shamon, "Secure spread spectrum water marking for multimedia", IEEE Tram. Proceed. 6, 12, 1673-1687, (1997)); Piva et al. By Blind SS (see A. Piva, M. Barni, E. Bartolini, V. Cappelini, “DCT-based watermarking recovering to the uncorrupted olIPol. 1, pp. 520); and Generalized Quantum Index Modulation (QIM) data concealment method (see B. Chen and GW Wornell, “Digital watermarking and using dither modulation,” Proceeding of IEEE MMSP 199 8, pp 273-278.IS (here, the step size is 5 and the embedding rate is 0.1 bpp)), and general-purpose LSB.

In these data hiding methods, various random or pseudo-random signals are embedded in various images. In general LSB data concealment, an embedding position was randomly selected for each image. Therefore, this method may be applied to a steganographic tool that uses LSB as a message embedding method. Various data embedding rates ranging from as low as 0.3 bpp to 0.01 bpp were applied. The range of this embedding rate is similar to that reported for the LSB-based stego tool in the above Lye and Farid. However, this evaluation may be considered more general due at least in part to the choice of the embedding location.

In this specific test evaluation, the threshold value 設定 was set to 4, but as described above, the claimed subject matter is not limited to a fixed threshold value or an integer value. The range of the effective prediction error value of the present embodiment is [−4 to 4], and has a total of nine different values. Therefore, the dimension of the transition matrix is 9 × 9, which is 81 features per error image. Since there are three error images in three different directions, the total number of features is 243 per image in this particular embodiment, but in this respect as well, the claimed subject matter is not limited in scope.

For the images in the image database, a stego image was generated by the data hiding method. Subsequently, the system was evaluated by the data hiding method. Half of the original set images were selected randomly or pseudo-randomly and used for training along with the corresponding stego images. To evaluate performance, in this embodiment, the remaining original and corresponding stego image pairs were subjected to a trained SVM. In this specification, the detection rate is defined as the ratio of the number of correctly classified images to the number of test images. The test procedure was performed 20 times. The following test data represents the average value of the number of tests.

First, a linear SVM process was applied. Linear SVM has the advantage of relatively fast training. However, similar benefits may not be provided for non-linear separable patterns. LIBSVM's Matlab SVM code was used (see the following literature: CC Change and CJ Lin, LIBSVM: a library for support machinery , 2001, http: // www. edu.tw/~cjlin/libsvm ). Table 1 shows the test results.

In Table 1, “TN” represents “true negative”, and is the detection rate of the original cover image in this specification. “TP” represents “true positive” and is a detection rate of a stego image in this specification. “Average” is the arithmetic average of these two detection rates. In other words, this is the overall correct classification rate of all test images.

Sullivan et al. Applied the same set of images and the same data hiding method. The same training and testing procedure was used. The results are listed in Table 2. According to this data, in the present example, the embodiment shown is based on Sullivan et al. Performance is better than

Similarly, in another evaluation, a polynomial kernel was used to train the 243D and 129D features obtained as described above. The results are listed in Table 3 and Table 4, respectively. As shown in the table, in this example, this particular embodiment has a true positive rate of 90% or more for Cox SS, Piva Blind SS, QIM, and LSB with embedded strength greater than 0.1 bpp. Have. Here, the embedded data includes an image having a size of 32 × 32 to 194 × 194. The corresponding embedded data rate is 0.02 bpp to 0.9 bpp, and the detection rate is 1.9% to 78%. Therefore, this particular embodiment appears to outperform the approach shown in Lyu and Farid when compared to the results reported in Lyu and Farid.

While specific embodiments have been described above, it will be appreciated that the claimed subject matter is not limited in scope to the specific embodiments or implementations. For example, in one embodiment, it may be implemented in hardware and may be implemented to operate, for example, in a device or combination of devices, and in another embodiment may be in software. Similarly, in certain embodiments, it may be implemented in firmware or as any combination of hardware, software, and / or firmware, for example. Similarly, in one embodiment, one or more articles may be included, such as storage medium (s), but claimed subject matter is not limited in scope in this respect. For example, these storage media, such as one or more CD-ROMs and / or disks, may have instructions stored thereon, such as executed by a system such as a computer system, computing platform or other system. As a result, an embodiment of the method according to the claimed subject matter may be performed, such as one of the above-described embodiments. As one possible example, a computing platform may include one or more processing units or processors, one or more input / output devices (such as a display, keyboard and / or mouse), and / or one or more memories ( Static random access memory, dynamic random access memory, flash memory, and / or hard drive, etc.). For example, a display may be used to display one or more queries (such as those with correlations) and / or one or more tree representations, again here the claimed subject matter The scope is not limited to this example.

In the foregoing description, various aspects of the claimed subject matter have been described. For purposes of explanation, specific numbers, systems and / or configurations are set forth in order to provide a thorough understanding of claimed subject matter. However, it will be apparent to one skilled in the art having the benefit of this disclosure that the claimed subject matter may be practiced without the specific details set forth above. In other instances, well-known features may be omitted and / or simplified so as not to obscure the claimed subject matter. Although specific features are illustrated and / or described herein, many variations, substitutions, modifications, and / or equivalents will occur to those skilled in the art. Therefore, it is to be understood that the appended claims are intended to cover all such modifications and / or modifications as fall within the spirit of the claimed subject matter.

1A to 1C are schematic diagrams illustrating an embodiment of a prediction error model applied to content such as an image.

Claims (33)

A method of training the content classification process,
Forming a plurality of prediction error sets from adjacent samples of the selected content;
Thresholding the formed prediction error set; and training the analysis of variance process using the thresholded prediction error set;
A method comprising processing the selected content. The content includes an image,
The method of claim 1. The analysis of variance processing includes SVM processing,
The method of claim 1. The plurality of prediction error sets includes at least three prediction error images;
The method of claim 1. The prediction error image includes a horizontal direction prediction error image, a vertical direction prediction error image, and a diagonal direction prediction error image.
The method of claim 4. The threshold processing includes non-uniform threshold processing,
The method of claim 1. A method for classifying content,
A method of applying a trained analysis of variance process to content and classifying the content based at least in part on values obtained by applying the trained analysis of variance process. The trained analysis of variance process includes a trained SVM process,
The method of claim 7. The content includes an image,
The method of claim 7. The trained analysis of variance process is based at least in part on a thresholded prediction error image;
The method of claim 9. A method of training the image classification process,
Forming three prediction error images from adjacent pixels of the selected image;
Thresholding the formed prediction error image; and training the SVM process using the thresholded prediction error image;
A method comprising processing the selected image. The prediction error image includes a horizontal direction prediction error image, a vertical direction prediction error image, and a diagonal direction prediction error image.
The method of claim 11. The threshold processing includes non-uniform threshold processing,
The method of claim 11. A method for classifying images,
A method of applying a trained SVM process to an image and classifying the image based at least in part on a value obtained by applying the trained SVM process. The trained SVM processing is based at least in part on a thresholded prediction error image;
The method according to claim 14. An article including a storage medium storing instructions,
As a result, when the instruction is executed,
Forming a plurality of prediction error sets from adjacent samples of the selected content;
Thresholding the formed prediction error set; and training the analysis of variance process using the thresholded prediction error set;
An article on which the selected content processing method is implemented. The content includes an image,
The article of claim 16. As a result, when the instruction is executed,
The analysis of variance processing includes SVM processing;
The article of claim 16. As a result, when the instruction is executed,
The plurality of prediction error sets includes at least three prediction error images;
The article of claim 16. As a result, when the instruction is executed,
The prediction error image includes a horizontal direction prediction error image, a vertical direction prediction error image, and a diagonal direction prediction error image,
The article of claim 19. As a result, when the instruction is executed,
The threshold processing includes non-uniform threshold processing;
The article of claim 16. An article including a storage medium storing instructions,
As a result, when the instruction is executed,
Applying a trained analysis of variance process to the content; and classifying the content based at least in part on values obtained by applying the trained analysis of variance process.
Articles on which the content classification method is implemented. As a result, when the instruction is executed,
The trained analysis of variance process includes a trained SVM process;
The article of claim 22. The content includes an image,
The article of claim 22. When the instruction is executed, as a result, the trained analysis of variance process is performed,
The instructions are based at least in part on a thresholded prediction error image;
25. The article of claim 24. Means for forming a plurality of prediction error sets from adjacent samples of the selected content;
Means for thresholding the formed prediction error set;
Means for training an analysis of variance process using the thresholded prediction error set. The content includes an image,
27. Apparatus according to claim 26. The means for training the analysis of variance process includes means for training the SVM process.
27. Apparatus according to claim 26. The means for thresholding includes means for non-uniform thresholding,
27. Apparatus according to claim 26. A means of applying trained analysis of variance processing to content;
Means for classifying the content based at least in part on values obtained by applying the trained analysis of variance process. Means for applying the trained analysis of variance process includes means for applying a trained SVM process;
The apparatus of claim 30. The content includes an image,
The apparatus of claim 30. The means for applying the trained SVM process is based at least in part on a thresholded prediction error image;
33. The apparatus according to claim 32.

Images