JP4610653B2 - Method for identifying marked images using statistical moments based at least in part on JPEG sequences - Google Patents

Description

The present application relates to classifying or identifying content such as, for example, marked images.

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 as not to obscure the claimed subject matter.

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.

Due to the widespread use of JPEG images, steganographic tools for JPEG images have increased in recent years, among which model-based steganography (MB), F5 and OutGuess are the latest. However, the development of new tools for identifying images containing hidden data remains desirable. According to the claimed subject matter, one embodiment described herein is at least a statistical moment derived at least partially from an image 2-D array and a JPEG two-dimensional array. Including partially based methods. In this particular embodiment, a first order histogram and / or a second order histogram may be used, but the claimed subject matter is not limited in scope in this respect. For example, in other embodiments, for example, higher order histograms may be utilized. However, in this particular embodiment, the moments of the two-dimensional characteristic function are also used from these histograms, but again, other embodiments are not limited in this respect. For example, a higher order moment may be used.

The widespread use of the Internet is accelerating due to the widespread use of computers. As a result, millions of videos are flowing on the Internet every day. In recent years, exchange of JPEG (Joint Photographic Experts Group) images has been performed more and more frequently. Many steganographic techniques for manipulating JPEG images have been released and are publicly used. Most of this type of technology appears to embed hidden data by changing 8x8 block discrete cosine transform (BDCT) coefficients in the JPEG domain. Among steganographic techniques, the recently published schemes OutGuess, F5 and model-based steganography (MB) appear to be the latest. See the following references: Provos, “Defending Against Statistical Steganization,” 10th USENIX Security Symposium, Washington DC, USA, 2001; Westfeld, "F5 a steganologic algorithm: High capacity de- fect better stegansis," 4th International Workshop on Information Hidden, Pittsburgh, USA Salee, “Model-based steganography,” International Work-shop on Digital Watermarking, Seoul, Korea, 2003. OutGuess embeds the data to be concealed using the redundancy of the cover image. Here, the cover image refers to content in which hidden data is not embedded. For JPEG images, OutGuess attempts to store statistics based at least in part on the BDCT histogram. In addition, OutGuess identifies redundant BDCT coefficients and embeds data in these coefficients to reduce the effects of data embedding. In addition, it tries to preserve the original BDCT histogram by adjusting the coefficients that are not embedded with data. F5, an extension of Jseg, F3 and F4, uses the following techniques: straddling and matrix coding. Stradling distributes the message as evenly as possible on the cover image. Matrix coding tends to improve embedding efficiency (defined herein as the number of embedding bits per change in BDCT coefficients). MB embedding attempts to correlate embedded data with a cover medium. This divides the cover medium into two parts, models the distribution parameters of the second part given the first part, and encodes the second part using the model and the embedded message Implemented by combining the two parts to form a stego medium. Specifically, a JPEG BDCT mode histogram is modeled using a Cauchy distribution, and embedding attempts to keep the low accuracy histogram of the BDCT mode unchanged.

Many steganalysis methods have been proposed for detecting hidden information in stego images. A generalized steganalysis method using higher order statistics has been proposed by Farid. See the following references: Farid, “Detecting hidden messages using high-order statistic-tick models,” International Conference on Image Processing, Rochester, NY, USA, 2002 (hereinafter “Farid”). Decompose the test image into wavelet subbands using orthogonal mirror filters. Higher order statistics are calculated from the wavelet coefficients of the high frequency subbands to form a group of features. Another group of features is similarly shown from the prediction error of the wavelet coefficients of the high frequency subbands. Literature: Y.M. Q. Shi, G .; Xuan, D.C. Zou, J. et al. Gao, C.I. Yang, Z .; Zhang, P.A. Chai, W .; Chen, C.I. Chen, "Steganalysis based on moments of characteristic functions using wavelet decomposition, prediction-error image, and neural network," International Conference on Multimedia and Expo, Amsterdam, Netherlands, 2005 (hereinafter referred to as "Shi et al.") Are described in The method uses the statistical moment of the characteristic function of the test image, its prediction error image, and their discrete wavelet transform (DWT) subbands as features.

However, Fridrich has proposed a steganalysis method designed specifically to deal with the JPEG steganographic scheme. See the following references: Fidrich, “Feature-based steganization for JPEG images and it's implications for future design, steganographic schemes,” 6th Information Hidden Corp. The method uses a set of selected features of relatively small size when detecting images with hidden data generated by Outguess, F5 and MB over other steganalysis methods as described above. Excellent performance. See the following literature: Kharrazi, H .; T.A. Sencar, N .; D. Memon, “Benchmarking steganographic and steganary techniques,” Security, Steganography, and Watermarking of Multimedia Content, 2005, San Jose, CA.

As suggested above, in contrast, one embodiment described herein is a method based at least in part on statistical moments derived at least in part from an image 2D array and a JPEG 2D array. I will provide a. In this particular embodiment, a first order histogram and / or a second order histogram may be used, but in other embodiments a higher order histogram may be used in addition, or depending on, for example, the application and many other possible factors, or It may be used instead. However, for this embodiment, the moment of the two-dimensional characteristic function is also used for steganalysis.

Here, steganalysis is considered as a task of two-class pattern recognition. Thus, for example, a test image is classified as either a stego image (one holding hidden data) or a non-stego image (one holding no hidden data).
The above references, Shi et al. Although a 78-dimensional feature vector is used in steganalysis as described in, the claimed subject matter is not limited in scope in this respect. However, again, Shi et al. The first half of the feature is generated based at least in part on a given test image and its 3-level Haar wavelet decomposition. The other half of the feature is based at least in part on the prediction error image and its three-level Haar wavelet decomposition. If the test image and the prediction error image are represented as LL ₀ subbands, 26 subbands are obtained. The discrete Fourier transform of the image histogram is referred to herein as a characteristic function (CF). Thus, Shi et al. In, the CFs of these subbands are calculated. The first three absolute moments of these CFs are used to form a 78-dimensional feature vector. The absolute moment is defined as follows:

(Where H (x _i ) is the CF component at frequency x _i and N is the total number of different value levels of the coefficients in the target subband). In this specification, the prediction error image is a difference between the original image and the prediction image. Prediction is described by (2) below and the 2 × 2 array shown in FIG.

This particular embodiment is quite different from this conventional approach. For example, as suggested above, this embodiment includes a statistical moment derived at least in part from a JPEG two-dimensional array associated with the image. For example, consider a two-dimensional array generated by applying an 8 × 8 block DCT to an image and quantizing using a JPEG quantization table. Note that these quantized JPEG BDCT coefficients are either positive, negative, or zero. Therefore, in this embodiment, absolute values should be applied to the coefficients, as shown at 205 in FIG. In this embodiment of the specification, the resulting two-dimensional array is referred to as a JPEG two-dimensional array.

Shi et al. Applying the above procedure referenced from to this JPEG 2D array yields another set of 78 features. This is explained in part by the block diagram 200 of FIG. 2 and is processed as applied to the frequency domain representation of the image (here DCT block representation) as illustrated. In this embodiment, the moment of CF is defined as in equation (1), where N is the total number of different absolute values of the JPEG quantized BDCT coefficients in the subject subband, but as claimed. The subject matter is not limited in scope in this respect. However, in this particular embodiment, computing a prediction error 2D array from a JPEG 2D array is not exactly the same as computing a prediction error 2D array for an image. For the zero element in the JPEG two-dimensional array, the prediction value for the zero element in the JPEG two-dimensional array is set as zero, as shown in Equation (3). Thus, in this embodiment, the zero DCT coefficient remains zero in the prediction error two-dimensional array. Further, as indicated by 215 in FIG. 2, an absolute value operation is applied to generate a two-dimensional prediction error array.

Most advanced steganographic schemes, such as OutGuess and MB, attempt to keep the changes in the histogram as small as possible so that they are difficult to identify or detect. For example, MB embedding attempts to keep the BDCT mode low accuracy histogram relatively unchanged. Thus, higher order histograms and moments may be useful in this situation.

A second order histogram may provide a measure of joint occurrence of a pair of pixels separated by a specific distance and orientation. Express the distance by ρ and the angle to the horizontal axis by θ. A secondary histogram is defined as follows:

(Where N (j ₁ , j ₁ ; ρ, θ) is the number of pixel pairs in which the first pixel value is j ₁ and the second pixel value is j ₂ , and N _T (ρ, θ) Is the total number of pixel pairs in the image with separation (ρ, θ)). A secondary histogram may be used in a JPEG two-dimensional array, but the claimed subject matter is not limited in scope in this respect. A quadratic histogram corresponds to a two-dimensional array often referred to as a dependency matrix or a co-occurrence matrix.

For wavelet subbands derived at least partially from a JPEG two-dimensional array, three secondary histograms with the following three separations are generated:

(These are referred to herein as a horizontal two-dimensional histogram, a vertical two-dimensional histogram, and a diagonal two-dimensional histogram, respectively). For example, referring to FIG. 1, the pairs (x, a), (x, b), (x, c) are (1, 0), (1, -π / 2), (1, -π / 4). After applying a two-dimensional DFT to the second order histogram to obtain a two-dimensional CF, two marginal moments of the two-dimensional CF may be calculated as follows:

(Where H (u _i , v _j ) is a two-dimensional CF component at frequency (u _i , v _j ) and N is the total number of absolute values with different coefficients in the target subband). For a certain direction, two limiting moments can be generated according to equations (6) and (7), so in this embodiment, 78 × 3 = 234 additional feature components are obtained. Thus, for this embodiment, as shown in FIG. 2, a 390 dimensional feature vector is obtained, but the claimed subject matter is not limited in scope in this respect.

Various techniques can be used to analyze data (referred to herein as features) in various situations. Here, the term “ANOVA” is used to refer to the following processes or techniques: As a result of applying the above processes or techniques, differences resulting from statistical fluctuations are differences resulting from non-statistical fluctuations. Processes or techniques that are sufficiently distinguished from each other to correlate, segment, classify, analyze, or otherwise characterize data based at least in part on the application of such processes or techniques. Examples include artificial intelligence technology and processing (such as pattern recognition), neutral networks, genetic processing, heuristics, support vector machines (SVM), etc., which are within the scope of the claimed subject matter. It is not limited.

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. 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 nonlinear 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 embodiment, for example, a linear kernel may be used for linear SVM processing. Similarly, other kernels may be used for nonlinear SVM processing.

Thus, Shi et al. Used a neural network, but in this embodiment, a support vector machine (SVM) is used as a classifier. In this particular embodiment, a polynomial kernel is used, as described below, but again the claimed subject matter is not limited in scope in this respect.

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.

An image database containing 7,560 JPEG images with quality factors of 70-90 was used. One third of the images were a set of substantially random photographs taken at different times and places with different digital cameras. The remaining two thirds were downloaded from the Internet. Each image was trimmed to either 768 x 512 or 512 x 768 size (leaving the central portion).

The above performance evaluation focuses on detection of OutGuess, F5 and MB1 steganography. Codes for these three approaches are publicly available. See the following document: http: // www. outgoings. org / ; http: // wwwrn. inf. tu-dressen. de / ~ westfeld / f5. html; http: // redwood. ucdavis. edu / phil / papers / iwdw03. htm . Since there are quite a number of zero BDCT coefficients in a JPEG image and the number of zero coefficients varies, the data embedding capacity varies from image to image. In general, the ratio of the length of hidden data to the number of non-zero BDCT AC coefficients is used as a measure of the data embedding capacity for JPEG images. For OutGuess, 0.05, 0.1 and 0.2 bpc (bits per non-zero BDCT AC coefficient) were embedded. As a result, 7498, 7452 and 7215 stego images were obtained, respectively. For F5 and MB1, 0.05, 0.1, 0.2, and 0.4 bpc are embedded, and 7560 stego images are obtained. Note that the MB1 embedded step size is equal to 2 in this evaluation.

Half of the images (and associated stego images) were randomly selected to train the SVM classifier and the remaining pairs were used to evaluate the trained classifier. Farid, Shi et al. The above-described method, such as the Fridrich method, and the above-described embodiment were applied to the detection evaluation of the OutGess, F5, and MB schemes. The results shown in Table 2 are arithmetic averages of 20 random tests.

In Table 1, the unit is%; OG represents OutGuess, TN represents the true negative rate, TP represents the true positive rate, and AR represents the accuracy. Please note that.

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 some 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 such as storage medium (s) may be included, 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.

FIG. 3 is a schematic diagram illustrating one embodiment of a 2 × 2 array of pixels. It is a block diagram showing one embodiment of feature generation.

Claims (37)

A method for classifying an image in a device that performs image processing, wherein features are generated based at least in part on a block frequency domain representation of the image , and the images are classified based at least in part on the generated features. There,
The method, wherein at least one of the features is based at least in part on a statistical moment that is generated based at least in part on at least one second order histogram of the block frequency domain representation of the image . The block frequency domain representation includes a DCT block representation;
The method of claim 1. The DCT block representation includes a two-dimensional JPEG array of the images;
The method of claim 2. The image is classified as either a stego image or a non-stego image,
The method of claim 1. Generating the feature includes generating a prediction image and a prediction error image based at least in part on the image;
The method of claim 1. Generating the feature includes generating a two-dimensional prediction error array based at least in part on the block frequency domain representation of the image;
The method of claim 1. The image classification is:
Applying trained classification means to the generated features;
The method of claim 1, wherein the trained classifier is trained based on features generated from a set of test images. 8. The method of claim 7, wherein the trained classifier is trained based at least in part on a statistical moment of a frequency domain representation of the set of test images. The frequency domain representation of the image includes a DCT block representation;
The method of claim 8 . The DCT block representation includes a JPEG representation,
The method of claim 9 . Computer
Means for generating features based at least in part on a block frequency domain representation of the image;
A computer readable recording medium having recorded thereon a program for functioning as means for classifying the image based at least in part on the generated feature,
A recording medium, wherein at least one of the features is based at least in part on a statistical moment generated based at least in part on at least one second order histogram of the block frequency domain representation of the image. The block frequency domain representation includes a DCT block representation;
The recording medium according to claim 11. The DCT block representation includes a two-dimensional JPEG array of the images;
The recording medium according to claim 12. The means for classifying the images is
Classified as either said image stego image or a non-stego image,
The recording medium according to claim 11. The means for generating the feature is:
Generating a prediction image and a prediction error image based at least in part on the image ;
The recording medium according to claim 11. The means for generating the feature is:
Generating a two-dimensional prediction error array based at least in part on the block frequency domain representation of the image ;
The recording medium according to claim 11. The means for classifying the images is
Applying trained classification means to the generated features;
The recording medium according to claim 11, wherein the trained classification means is trained on the basis of features generated from a set of test images. 18. The recording medium of claim 17, wherein the trained classification means is trained based at least in part on a statistical moment of a frequency domain representation of the set of test images. The frequency domain representation of the image includes a DCT block representation;
The recording medium according to claim 18. Before Symbol DCT block representation, including the JPEG representation,
The recording medium according to claim 19. Means for generating features for classification based at least in part on a block frequency domain representation of the image ;
Means for classifying the image based at least in part on the generated classification features , comprising:
The apparatus characterized in that at least one of the features is based at least in part on a statistical moment generated based at least in part on at least one second order histogram of the block frequency domain representation of the image . The block frequency domain representation includes a DCT block representation;
The apparatus of claim 21. The DCT block representation includes a two-dimensional JPEG array of the images;
The apparatus of claim 22. The means for classifying includes means for classifying an image as either a stego image or a non-stego image.
The apparatus of claim 21. The means for generating the feature includes means for generating a prediction image and a prediction error image based at least in part on the image.
The apparatus of claim 21. The means for generating the features includes means for generating a two-dimensional prediction error array based at least in part on the block frequency domain representation of the image.
The apparatus of claim 21. Means for applying a trained set to the means for classifying;
The apparatus of claim 21, wherein the trained set includes a set of features for classification generated based on a set of trained images. 28. The apparatus of claim 27, wherein the trained set includes at least one statistical moment of a frequency domain representation of the trained image set. The frequency domain representation of the image includes a DCT block representation;
30. The apparatus of claim 28. The DCT block representation includes a JPEG representation,
30. Apparatus according to claim 29. An apparatus for analyzing images,
A frequency domain transform that processes the image and provides a block frequency domain representation of the image;
At least one statistical moment generator that generates a first set of image features based at least in part on the block frequency domain representation;
A classifier for classifying the image based at least in part on the first set of image features;
The apparatus, wherein the first set of image features includes a statistical moment generated based at least in part on at least one second order histogram of the block frequency domain representation of the image. 32. The apparatus of claim 31, comprising:
Comprising at least one histogram forming unit for obtaining at least one frequency domain histogram based on the block frequency domain representation;
The apparatus, wherein at least one of the statistical moment generators forms the first set based on at least one frequency domain histogram. An apparatus according to claim 32, comprising:
The apparatus, wherein the at least one histogram forming unit provides at least one frequency domain histogram selected from the group consisting of a horizontal two-dimensional histogram, a vertical two-dimensional histogram, and a diagonal two-dimensional histogram. 32. The apparatus of claim 31, comprising:
An error prediction unit that acquires a two-dimensional prediction error array based on the block frequency domain representation,
The apparatus wherein the at least one statistical moment generator forms a first set of image features based on the prediction error two-dimensional array based on the block frequency domain representation. 32. The apparatus of claim 31, comprising:
A histogram forming unit that acquires an image histogram based on a spatial representation of the image;
At least one statistical moment generator generates a second set of image features based on the image histogram;
The apparatus, wherein the classifier further classifies the image based on a second set of image features. 36. The apparatus of claim 35, comprising:
An error prediction unit that obtains a two-dimensional prediction error array based on a spatial representation of an image;
The apparatus, wherein at least one statistical moment generator generates a second set of at least some image features based on the two-dimensional prediction error array based on a spatial representation of the image. 32. The apparatus of claim 31, comprising:
The device is
Operating the frequency domain transform and at least one statistical moment generator to generate at least one trained feature set for the trained image;
An apparatus for training the classifier based on at least one set of trained features.

Images