Научные стремления 2012-1

УДК 004.056

Demidchuk Aleksey1

STEGANALYSIS OF JPEG CONTAINERS

1Belarusian State University of Informatics and Radioelectronics, Minsk

I.INTRODUCTION

Digital steganography combines secure communication methods, which use digital objects as containers. The most commonly as containers are used images, video, audio data. In forming a digital signature for intellectual property objects some data is being embedded so, that information is stored in different modifications of the object. Also methods of steganography can be applied for concealed copying of secret data [1].

Jian Zhao and Eckhard Koch offered to perform inlining of covered message with the JPEG compression [2,3]. Embedding is performed in coefficients of discrete cosine transform (DCT) that is grouped in 8x8 blocks. For that the three coefficients are selected pseudo-randomly from medium frequency domain which highlighted in fig. 1. These coefficients are modified as follows: for embedding bit '1', the first two coefficients should be more than a third with a value D; for embedding bit '0' the first two coefficients should be less than a third coefficient with a value D. Other combinations are not considered valid.

Figure 1.Embedding area

II.STEGANALYSIS

Steganalysis method is based on calculation of the standard deviation (STD) for the coefficients of DCT and analysis on the distribution histogram of the coefficients of STD for whole image.

Image steganalysis has two parts:

test case embedded with a threshold value D > 1; test case embedded with a threshold value D = 1.

III.TEST CASE EMBEDDED WITH A THRESHOLD VALUE D > 1

Of each block of DCT coefficients are selected 8 coefficients (highlighted in fig. 1) -- denote block b8. Calculate STD for every block b8:

414

_

where z = 8 – count of selected DCT coefficients; xj – j-th coefficient; x – the average ratio in the test group.

The histogram HSTD(S) is calculated for the s>0 coefficients. Example histograms for source image and stego container with embedding process which was using threshold value D=2 are shown in Figure 2: a, b, respectively.

a) Source image

b) Stego image with D=2

Figure 2.Examples of histograms of STD coefficients

Maximum of the histogram shows the most frequent STD value:

This STD value is compared with threshold value q: smax ≤ q => Image {STEGO},

smax> q => Image {STEGO}.

If smax> q then the analyzed component of image contains hidden information, otherwise this is empty container. Threshold value q = 0.354.

IV. TEST CASE EMBEDDED WITH A THRESHOLD VALUE D = 1

For images that may contain hidden information with threshold D=1, the following value is calculated:

415

count of STD coefficients. The histograms, that shown on Figure 3, was built for the 400 images: with D=1 threshold (line ID = 1) and with no embedded information (line I0). Figure 2 shows that if value PHmax> 35, then image contains hidden information

with 100% probability. But if PHmax ≤ 35, then the probability of the existing hidden information content can be got from the Figure 4. This curve was obtained using the

following formula for the histograms from Figure 3:

Figure 3.Histogram of the distribution of values PHmax for the sets I0 and ID = 1

Figure 4.The dependence on the value of PHmax of the probability of occurrence hidden information in the image

For example, image, that have PHmax= 25, contains hidden information with probability PS= 0.65.

V.RESULTS

418 images with different resolution from 384 × 288 до 3089 × 2084 were used as a test set. Containers were filled to 100% with threshold values D = {1, 2, 3}. Table. 1 shows the results of the estimation of information embedded on the test set

416

of images using the OutGuess and JPHide steganalysis methods, which is included in Stegetect program [4].

Table 1.The results compare the effectiveness analysis, %

Where e1 – error of the first kind, i.e. estimation ratio of false positives for the empty image, to the total number of empty images; e2 – error of the second kind, i.e. estimation ratio of missed filled images as empty containers to the total number of images.

VI. CONCLUSION

In summary, using the proposed criteria for estimating the JPEG images to determine the presence of hidden information by the method of Koch – Zhao, under certain conditions, provides effective solutions of steganalysis. Compared to existing methods, oriented to other embedding algorithms, greatly increases the accuracy of the analysis.

REFERENCES

[1]E. Cole. ―Hiding in Plain Sight: Steganography and the Art of Covert Communication‖.

Wiley Publishing, Inc. 2003.

[2]J. Zhao, E. Koch, Embedding Robust Labels into Images for Copyright Protection. Proc. of the Int. Congress on Intellectual Property Rights for Specialized Information, Knowledge and New Technologies. – Munich, Vienna, 1995. – P. 242–251.

[3]J. Zhao, E. Koch, Towards Robust and Hidden Image Copyright Labeling. IEEE Workshop on Nonlinear Signal and Image Processing. – Greece, 1995. – P. 123–132.

[4]Ошибка! Недопустимый объект гиперссылки. [Electronic resource]. – Mode of access

:http://www.outguess.org. – Date of access : 11.12.10.

Demidchuk A.I. 1

Steganalysis of JPEG containers

1 Belarusian State University of Informatics and Radioelectronics.

Summary

The article provides a brief description of the steganographic data hiding technique by Koch

– Zhao. It describes the algorithm of analysis of images in JPEG format to establish the fact hidden communication, and the results of application to a test set of images. The result enables us to offer this method for primary selection of JPEG images for subsequent detailed analysis.

Number of equations: 4. Number of Pictures: 7. Number of references: 11.

417

УДК 004.932.1

Yanochkin A.L.

Localization of symbols zones on digital images and video

Belarusian State University of Informatics and Radioelectronics

I. INTRODUCTION

An important component in the development of automatic text recognition system is splitting a input image to individual characters zones. The overall recognition rate of the system strongly depends on the quality of this step. There are many factors that cause the character segmentation task difficult, such as image noise, plate frame, rivet, space mark, plate rotation and illumination variance. Compared to traditional optical character recognition (OCR), the text images captured using video cameras contain objects of interest with complex backgrounds. It is a challenge for researchers to detect objects of interest from a complex and low contrast background. The input data is a sequence of images or a set of video frames presented as a sequence of images. The initial parameters of the system are maximum angle of rotation of text line, a rotation angle delta. Our goal is to separate character zones.

II. IMAGE PRE-PROCESSING

A sequence of color images is an input of the system. First at all the image or the first frame is transferred to grayscale by the standard formula (1).

G = 0.299*R + 0.587*G + 0.114*B. (1)

Fig. 1. Original input image

Fig. 2. Grayscale image.

The following step of the algorithm is a binarization by adaptive thresholding since the license plate is not evenly lighted. Adaptive thresholding takes a grayscale image as input and outputs a binary image. For each pixel in the image, a threshold has to be calculated. If the pixel value is below the threshold it is set to the background value, otherwise it assumes the foreground value.[1]

III. DETERMINATION OF THE TEXT ANGLE

To determine the correct orientation of the license plate on the image we generate a set of rotated by a delta angle images and estimate each of them for the best

418

approximation of the horizontal orientation. The calculated horizontal and vertical projection of each of the intermediate images are robust enough for solving the task.

Fig. 3. Thresholding image.

The vertical and horizontal projections are to be smoothed for the correct calculating the local maxima and minima values. At an optimal projection we can get a minim minimum amounts and a maximum of a difference between neighboring local extremums.

Fig. 4. The vertical projection of the image

Fig. 5. The rotated image for a 3 degrees angle

Fig. 6. The vertical projection of the rotated image for an angle of 3 degrees

Then image is divided into parts by vertical and a horizontal line in accordance of local extremums of the projections [9,11]. Areas what do not fit to the minimum character sizes are discarded. If appropriate parts are not found (for example when after thresholding merged letters and background) the next image is taken from an image sequence and the whole algorithm starts from the beginning. We check the correctness of the obtained angle values for the next image of the sequence. If both

419

images are approximately at the same angle so the angle is obtained. And two images are rotated for this angle. In case of the angles are different then the last image is taken as the base and the algorithm starts from the beginning again. If the algorithm couldn‘t get the same angle for any pair of the images of the sequence, so it is considered that the sequence is not acceptable and the algorithm stops.

IV. SEGMENTING SYMBOLS

For the ―multiline‖ plates we have to segment symbols of every text line separately.

On the basis of the vertical projection the input image is split to parts corresponding of every line, in case we need it. For every part of the initial image the horizontal distance between the characters are defined. If the distance between the characters is about the same, then we consider that we have a plate on a base of a proportional font, in other case there is a monospaced font. The monospaced font have, each character of a same fixed width. In proportional font the same distance (spaces) between the characters are used.[6] After this the string is divided into individual characters. If the character does not fit the proportions, for example, wider than the widest character (merger characters) then the partition is considering as a wrong and we have to return to the previous step of the algorithm (i.e. to the previously defined angle and to divide into symbols according to certain characteristics of the font). If this image does not fit the characters well, this picture is taken at the base and the partitioning algorithm on the characters is repeated.

Fig. 7. Examples of proportional font and monospaced font

If there was not found any suitable image in the entire sequence, you should select best of images, in which the smallest number of characters does not match the proportions. In this picture, the inscription is divided according to previously defined font characteristics for this image.

REFERENCES

1.Gonzalez, Rafael C. & Woods, Richard E. Thresholding. In Digital Image Processing, pp. 595–611, 2002

2.Shapiro, Linda G. & Stockman, George C. "Computer Vision". Prentice Hall, 2002

3.Milan Sonka, Vaclav Hlavac and Roger Boyle. Image Processing, Analysis, and Machine Vision. Thomson, 2008

4.Turek, Fred. "Machine Vision Fundamentals, How to Make Robots See", pp. 60-62 2011

5.E. Roy Davies. Machine Vision: Theory, Algorithms, Practicalities. Morgan Kaufmann, 2005.

6.Headley, Gwyn. The Encyclopaedia of Fonts. Cassell Illustrated: 2005

7.Thomas, George B.; Weir, Maurice D.; Hass, Joel. Thomas' Calculus: Early Transcendentals. 12th ed., 2010.

420