Retaliating Anti-forensics of JPEG Image Compression Based

On the Noise Level Estimation

INTERIM PRESENTATIONSPRING 2015

ADVISOR: Dr. K.R.Rao

Presented by, Komandla Sai Venkat, UTA id: 1001115386

saivenkat.komandla@mavs.uta.edu

Acronyms

• JPEG: The Joint Photographic Experts Group• DCT: Discrete Cosine Transform• FPR: False positive rates • TPR: True positive rates • TNR: True negative rates• TIFF: Tagged Image File Format• ROC: Receiver operating characteristics • UCID: Uncompressed Colour Image Database

TABLE OF CONTENTS

ObjectiveJPEG compression and its ant-iforensicsAnti-forensic dither additionSummarizing de-blocking theoremBlocking detection The noise level detectionPerformance characteristics Experimental resultsReferences

OBJECTIVE

• To implement various steps involved in retaliation of the anti-forensics of the jpeg compressed image based on the noise level estimation .

• To estimate the noise level of a particular image and to compare it with a threshold to determine whether it is forged.

JPEG compression and its antiforensics

JPEG (Joint Photographic Experts Group) is very well known ISO/ITU-T standard created in late 1980's.JPEG compression can be divided into four steps: colour mode conversion and sampling, DCT transform, quantization, and entropy coding.JPEG compression starts by segmenting an input image into several non overlapping 8 × 8 pixel blocks, then it uses the 2-D DCT to transform each block data into 64 DCT coefficients.

Block Diagram of JPEG Compression [7]

ANTI-FORENSICS OF JPEG COMPRESSION

• An anti-forensic method [6] can deceive detectors by first adding anti-forensic dither to DCT transform coefficients to imitate the original uncompressed histograms and then erasing blocking artifacts to remove the compression history by boundary blurring.

• To erase the compression history, the forger must remove blocking artifact by first median filtering an image and then adding low-power white Gaussian noise to each of its pixel values.

• In order to detect the forged images, noise level estimation [7] has been employed to estimate the noise added in the deblocking process.

Anti-forensic modified process of JPEG compression [6]

Anti-forensic dither

• In order to remove the quantization artifact for a JPEG compressed image, i.e. to make the sub band coefficient value distribution match an original one, the anti forensic dither is added to the DCT coefficient.

• To hide the compression evidence, [6] dither is introduced into the AC coefficients to approximately restore the histogram of each subband, by:

Z = Y + D , where Z is the anti-forensically modified coefficient and D is the additive dither.

ORIGINAL IMAGE

JPEG compressed image using a quality factor of 90(a), 70 (b), 30 (C), and 10 (d) followed by the addition of anti-forensic dither to its DCT coefficients.

THE NOISE DISTRIBUTION

The noise distribution for the coefficient Y of zeros value at the (i, j)-th position, is given by:

DISTRUBUTION OF ANTIFORENSIC DITHER

• The distribution of the anti forensic dither added to nonzero quantized DCT coefficients is given by:

Quality factor

• Image quality is the measure of how accurately our image matches the source image which is observed by visible factors like brightness and evenness of illumination, contrast, resolution, geometry, colour fidelity and colour discrimination of an observed image.

• Most implementations of JPEG compression use a set of quantization matrices indexed by a quality factor from the set {1, 2, . . . , 100} which are used in the reference implementation provided by the Independent JPEG Group. [19].

• A parameter called Q factor IS used to “tune” the quality of the JPEG image which vary from range between 1to 100.

• A factor 1 produces the image with maximum compression (i.e. smallest) but with worst quality .

• The factor of 100 produces the image with least compression (i.e. largest) but best quality.

JPEG compressed image using a quality factor of 65

HISTOGRAMS OF (2, 2) DCT COEFFICIENTS TAKEN FROM AN UNCOMPRESSED VERSION OF THE IMAGE (A), JPEG COMPRESSION

OF THE SAME IMAGE(B)

ANTI-FORENSICALLY MODIFIED IMAGE

HISTOGRAM OF AN ANTI-FORENSICALLY MODIFIED COPY OF THE JPEG COMPRESSED IMAGE

DIFFERENCE WITHIN A BLOCK AND SPANNING ACROSS A BLOCK BOUNDARY

HIGHWAY_CIF(352*288)

FOR EACH BLOCK (I, J), THE NUMBERS :

• Z'(i, j) = |A + D - B - C| and• Z’’ (i, j) = |E + H - F - G| are computed.• By examining the difference between two histograms,

the blocking artifacts are detected:

Where h1 is the histogram of Z' each image block, and h2 is the histogram of Z''.

For an uncompressed image, h1 (1) > h1 (0) and h2 (1) > h2 (0) and for a JPEG compressed image, h1 (1) > h1 (0) and h2 (1) > h2 (0) may not meet or not meet at the same time.

H1 AND H2 OBTAINED FROM AN UNCOMPRESSED IMAGE THE SAME IMAGE AFTER JPEG COMPRESSION USING A QUALITY FACTOR OF 75

H1 AND H2 OBTAINED FROM THE SAME IMAGE AFTER JPEG COMPRESSION USING A QUALITY FACTOR OF 75

MEDIAN FILTERING

• Difference between h1 and h2 is removed to remove blocking artifacts:

ADDING NOISE AFTER MEDIAN FILTERING

SUMMARY OF THE DE-BLOCKING ALGORITHM

Where ui,j represents the pixel value at location (i,j)

in an unmodified image, and vi,j denotes its de-blocked counterpart, meds( ) denotes a two-dimensional median filter with a square window of size s pixels, and ni,j is a zero mean Gaussian random noise with variance .

BLOCKING ARTIFACT DETECTION ACCURACY

RESULTSResults of blocking artifact detection accuracy from experiments by varying the window size(s) and the variance σ for FPR (False positive rates) varies in a given interval and the optimal threshold to obtain the Accuracy Rate = (TPR + TNR) / 2. TPR is true positive rates and TNR is true negative rates. The results show that the above method can remove the blocking artifact effectively.

THE NOISE LEVEL ESTIMATION

PERFORMANCE OF THE NOISE LEVEL ESTIMATION USING THE TWO METHODS

SUMMARY ON NOISE LEVEL ESTIMATION

• The noise level estimation can be summarized as the following major steps:• 1. Decomposing the test image into overlapping patches. The default patch

size is 7 × 7 pixels.• 2. Estimating an initial noise level σe from the covariance matrix as :•

• Where Σ y' is the covariance matrix of the selected patches and is the minimum Eigen value of Σ y'.• 3. Selecting the weak textured patches from the test image using a threshold

that varies with σe..

• 4. Estimating a new noise level σe using the selected patches. The process of step 3 and 4 is iterated until σe is stable. Here σe is used to denote the estimated noise level of a test image.

EXPERIMENTAL RESULTSROC curves with proposed method for different

quality factors:

DETECTION ACCURACY WITH DIFFERENT QUALITY FACTORS AND DIFFERENT PARAMETERS

(PROPOSED METHOD)

Proposed countering anti-forensic method can achieve average detection accuracy above 95%.

COMPARING THE PERFORMANCE BETWEEN THE PROPOSED METHOD AND OTHER METHODS

The results shown in the table compare the performance between the proposed method and other methods. All results are obtained with the median filter window size s = 3 and noise standard variance σ = 2.D represents the dimension and QF represents the quality factor.

CONCLUSIONS• A retaliation method has been proposed applied to

anti-JPEG compression method and implemented based on noise level estimation by assuming that the forger adopts the anti-forensic method proposed in [6] and achieved the ROC curves for different quality factors.

• Though anti-forensic technique can remove JPEG compression trace, it also introduces other detectable artifacts.

• In this presentation, we limited our analysis to only 250 images out of 1338 images from the Uncompressed Color Image Database (UCID) [12].

REFERENCES[1] J. Lukas and J. Fridrich, “Estimation of primary quantization matrix in double

compressed JPEG images,” in Proc. of SPIE 6819,Security, Forensics, Steganography, and Watermarking of Multimedia Contents X, pp. 681911, February 26, 2008.

[2] T. Pevny and J. Fridrich, “Detection of double-compression in JPEG images for applications in steganography,” IEEE Transactions on Information Forensics and Security, vol. 3,no.2, pp. 247-258, June 2008.

[3] W. Luo, Z. Qu and J. Huang,G. Qiu, “A novel method for detecting cropped and recompressed image block,” in Proc. of IEEE International Conference on Acoustics, Speech and Signal Processing 2007 (ICASSP 2007), pp. II-217-II-220, Feb. 2003.

[4] Z. Fan and R. L. de Queiroz “Identification of bitmap compression history: JPEG detection and quantizer estimation,” IEEE Transactions on Image Processing, vol.12, no. 2, pp. 230-235,Feb. 2003.

[5] D. Fu, Y. Q. Shi, W. Su, “A generalized Benford’s law for JPEG coefficients and its applications in image forensics,” In Proc. of SPIE-IS&T Electronic Imaging, SPIE Vol. 6505,Security, Steganography, and Watermarking of Multimedia Contents, pp. 65051L-165051L-11, Jan. 2007, San Jose, CA,USA.

REFERENCES

[6] M. C. Stamm and K. J. R. Liu, “Anti-forensics of digital image compression,” IEEE Transactions on Information Forensics and Security, vol. 6, no. 3, pp. 1050-1065, Sept. 2011.

[7 ]X. Liu, M. Tanaka and M. Okutomi. “Noise level estimation using weak textured patches of a single noisy image,” in Proc. 19th IEEE International Conference on Image Processing 2012(ICIP2012), pp. 665-668, IEEE, June 2012.

[8] E Y Lam and J W Goodman, "A mathematical analysis of the DCT coefficient distributions for images,” IEEE Transactions on Image Processing , vol. 9, no. 10, pp. 1661-1666,Sept. 2000.

[9] L. Liu and X. Zhuang, “A novel square root rate control algorithm for H. 264/AVC encoding, ” In Proc. of IEEE International Conference on Multimedia and Expo 2009( ICME 2009), pp. 814-817, Dec.2009.

[10] L. Liu, X. Zhuang, Z. He, and Y. Sun. “H. 264/AVC rate control with enhanced rate-quantisation model and bit allocation, ” IET image processing, vol. 5, no. 7, pp. 619-629,Sept.2011.

REFERENCES[11] D Zoran and Y. Weiss “Scale invariance and noise in natural images,” IEEE 12th

International Conference on Computer Vision 2009, pp. 2209-2216, IEEE,April 2009.[12] G. Schaefer and M. Stich, “UCID-An uncompressed color image database,” in Proc.

SPIE 5307, Storage and Retrieval Methods and Applicat. for Multimedia, pp. 472–480,Jan. 2004.

[13] G. Valenzise, V. Nobile and M. Tagliasacchi, et al. “Countering JPEG anti-forensics,” in Proc. 18th. IEEE International Conference on Image Processing 2011 (ICIP2011), pp. 1949-1952, IEEE, Dec. 2011.

[14] H. Li, W. Luo and J. Huang, “Countering anti-JPEG compression forensics,” in Proc. 19th. IEEE International Conference on Image Processing 2012 (ICIP2012), pp. 241-244, IEEE, Jan. 2012.

[15]M. C. Stamm, W. S. Lin and K. J. R. Liu “Forensics vs. antiforensics:A decision and game theoretic framework,”IEEE International Conference on Acoustics, Speech and Signal Processing2012 (ICASSP2012), pp. 1749-1752,

IEEE,Dec. 2012.

REFERENCES

[16]A.Nosratinia, "Enhancement of JPEG-Compressed images by re-application of JPEG," Journal of VLSI Signal Processing, vol. 27, pp. 69-79, 2001.

[17] W. B. Pennebaker and J. L. Mitchell. JPEG - Still Image Data Compression Standard. Van Nostrand ReinHold, NY,Sept. 1993.

[18] E. Hamilton,JPEG File Interchange Format - Version 1.02. C-Cube Microsystems, Sept. 1992.