JPEG Anti-forensics Using Non-parametric DCT Quantization Noise - PowerPoint PPT Presentation

JPEG Anti-forensics Using Non-parametric DCT Quantization Noise Estimation and Natural Image Statistics The 1st ACM Workshop on Information Hiding and Multimedia Security Wei Fan *, † , Kai Wang * , Fran¸ cois Cayre * , and Zhang Xiong † * GIPSA-lab, France † Beihang University, P. R. China 18-06-2013

Introduction Reliability of detectors DCT histogram smoothing Anti-forensics Conclusions What is forensics/anti-forensics to JPEG compression? Forensics The technique detecting tampering in digital media Anti-forensics The technique misleading forensic analyses Question : Has it been previously JPEG compressed? JPEG Anti-forensics Using Non-parametric DCT Quantization Noise Estimation and Natural Image Statistics 2 / 24

Introduction Reliability of detectors DCT histogram smoothing Anti-forensics Conclusions State-of-the-art 0.2 0.18 0.16 0.14 0.12 0.1 0.08 0.06 0.02 0.04 0.018 0.02 0.016 0 0.014 −100 −80 −60 −40 −20 0 20 40 60 80 100 DCT coefficient frequency 0.012 Quantization table 0.01 0.008 0.006 TV-based 0.004 0.002 estimation 0 −100 −80 −60 −40 DCT coefficient value −20 0 20 40 60 80 100 detector DCT histogram smoothing Standard JPEG + compression Median filtering based deblocking Calibration-based detector JPEG blocking detection process Countering JPEG forensics JPEG anti-forensics JPEG anti-forensics - Z. Fan and R. L. De Queiroz, “Identification of bitmap compression history: JPEG detection and quantizer estimation,” ITIP , 2003 - M. Stamm et al., “Anti-forensics of JPEG compression,” ICASSP , 2010 - M. Stamm et al., “Undetectable image tampering through JPEG compression anti-forensics,” ICIP , 2010 - G. Valenzise et al., “Countering JPEG anti-forensics,” ICIP , 2011 - S. Lai and R. B¨ ohme, “Countering counter-forensics: the case of JPEG compression,” IH , 2011 JPEG Anti-forensics Using Non-parametric DCT Quantization Noise Estimation and Natural Image Statistics 3 / 24

Introduction Reliability of detectors DCT histogram smoothing Anti-forensics Conclusions Our objective: performing better JPEG anti-forensics • Better undetectability when tested under current existing JPEG forensic detectors ✓ Quantization table estimation based detector D F q ✓ JPEG blocking artifacts detector D F b ✓ Total variation based detector D V ✓ Calibration based detector D L • Higher image visual quality of processed images compared with state-of-the-art methods ✓ F S q , JPEG image processed by DCT histogram smoothing ✓ F S q S b , JPEG image processed by DCT histogram smoothing followed by median filtering based deblocking JPEG Anti-forensics Using Non-parametric DCT Quantization Noise Estimation and Natural Image Statistics 4 / 24

Introduction Reliability of detectors DCT histogram smoothing Anti-forensics Conclusions JPEG anti-forensic experiment setup • 1338 genuine, uncompressed images of size 512 × 384 from UCID corpus (v2) are used for test • Without loss of generality, only the luminance component of the image is considered • UCIDTest contains the first 1000 images, while UCIDTrain contains the last 338 images • Each UCIDTest image I is compressed with a random quality factor Q ∈ [30 , 35 , 40 , 45 , 50 , 55 , 60 , 65 , 70 , 75 , 80 , 85 , 90] to JPEG image J • All types of JPEG anti-forensic forgeries are created from J JPEG Anti-forensics Using Non-parametric DCT Quantization Noise Estimation and Natural Image Statistics 5 / 24

Introduction Reliability of detectors DCT histogram smoothing Anti-forensics Conclusions Quantization table estimation based detector Tested on BOSSBase (v1.01), which consists of 10,000 genuine, uncompressed images of size 512 × 512 • F FP : false positive rate • P 3 : the portion of 3 among all “non-1”/“non-undetermined” estimated quantization table entries Archive P F P (%) P 3 01 18 . 70 1653 / 1959 = 84 . 38% 02 12 . 40 957 / 1118 = 85 . 60% 03 4 . 50 175 / 187 = 93 . 58% • Tested on UCIDTest 04 5 . 70 111 / 125 = 88 . 80% • P FP = 16 / 1000 = 1 . 6% 05 10 . 50 986 / 1141 = 86 . 42% 06 61 . 50 9243 / 11033 = 83 . 78% • P 3 = 15 / 16 = 94 . 74% 07 17 . 50 1084 / 1155 = 93 . 85% 08 43 . 90 5655 / 6670 = 84 . 78% 09 33 . 20 5062 / 5895 = 85 . 87% 10 44 . 30 6780 / 7720 = 87 . 82% Average 25 . 22 87 . 49% JPEG Anti-forensics Using Non-parametric DCT Quantization Noise Estimation and Natural Image Statistics 6 / 24

Introduction Reliability of detectors DCT histogram smoothing Anti-forensics Conclusions Reliability analysis 0.1 0.09 0.6 0.08 DCT coefficient frequency 0.5 DCT coefficient frequency 0.07 0.06 0.4 0.05 0.3 0.04 0.03 0.2 0.02 0.1 0.01 0 0 −10 −8 −6 −4 −2 0 2 4 6 8 10 −50 −40 −30 −20 −10 0 10 20 30 40 50 DCT coefficient value DCT coefficient value When does the detector pick 3 instead of 1 ? The likelihood for q = 3 : number of blocks in estimation ���� L (3) = p 0 × N × w (0 , 3) + (1 − p 0 ) × N × w (1 , 3) + N × log 3 � �� � The likelihood for q = 1 : an even function defined in the estimation L (1) = N × w (0 , 1) + N × log 1 L (3) > L (1) , when: percentage of coefficients which are integer multiples of 3 ���� p 0 > 67 . 28% JPEG Anti-forensics Using Non-parametric DCT Quantization Noise Estimation and Natural Image Statistics 7 / 24

Introduction Reliability of detectors DCT histogram smoothing Anti-forensics Conclusions The other three forensic detectors 1 1 0.9 0.9 0.8 0.8 Random guess 0.7 0.7 True positive rate True positive rate 0.6 0.6 0.5 0.5 0.4 0.4 D F b D F b 0.3 0.3 0.2 0.2 D V D V 0.1 0.1 D L D L Random guess 0 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 False positive rate False positive rate F S q F S q S b • The cost for F S q S b : to lose 5 . 13 dB of PSNR value compared to JPEG images on average JPEG Anti-forensics Using Non-parametric DCT Quantization Noise Estimation and Natural Image Statistics 8 / 24

Introduction Reliability of detectors DCT histogram smoothing Anti-forensics Conclusions Stamm et al.’s method ( F S q ) The key assumption The unquantized DCT coefficients follow the Laplacian distribution for AC components 0.16 0.045 0.04 0.14 0.035 0.12 0.03 DCT coefficient frequency DCT coefficient frequency 0.1 0.025 0.08 0.02 0.06 0.015 0.04 0.01 0.02 0.005 0 0 −250 −200 −150 −100 −50 0 50 100 150 200 250 −250 −200 −150 −100 −50 0 50 100 150 200 250 DCT coefficient value DCT coefficient value F S q original JPEG Anti-forensics Using Non-parametric DCT Quantization Noise Estimation and Natural Image Statistics 9 / 24

Introduction Reliability of detectors DCT histogram smoothing Anti-forensics Conclusions The addition of the dithering signal Strategy Noise is added randomly without any consideration of the local information of the image in the spatial domain F S q (log + scale) original (log + scale) JPEG Anti-forensics Using Non-parametric DCT Quantization Noise Estimation and Natural Image Statistics 10 / 24

Introduction Reliability of detectors DCT histogram smoothing Anti-forensics Conclusions JPEG image restoration Why? • Improve the image quality for the further loss during the forgery creation • Improve the accuracy of the DCT quantization noise estimation using calibration Assumptions • x q = x + n q : JPEG image x q is obtained by adding spatial-domain quantization noise n q to the original image x • n q is a random quantity, n q and x are independent MAP criterion ˆ x = arg max p ( x | x q ) = arg max p ( x q | x ) p ( x ) = arg max p ( n q ) p ( x ) x x x - M. A. Robertson and R. L. Stevenson, “DCT quantization noise in compressed images,” ITCSVT , 2005 JPEG Anti-forensics Using Non-parametric DCT Quantization Noise Estimation and Natural Image Statistics 11 / 24

Introduction Reliability of detectors DCT histogram smoothing Anti-forensics Conclusions Using the EPLL framework By maximizing the EPLL (Expected Patch Log Likelihood), we expect to find ˆ x in which every patch is likely under the patch prior • The quantization noise model: 0 -mean multivariate Gaussian • The patch prior model: GMM (Gaussian Mixture Model) • 64 kinds of patches according to their relative position w.r.t. JPEG blocks • The learning of models is performed on UCIDTrain - D. Zoran and Y. Weiss, “From learning models of natural image patches to whole image restoration,” ICCV , 2011 JPEG Anti-forensics Using Non-parametric DCT Quantization Noise Estimation and Natural Image Statistics 12 / 24

Introduction Reliability of detectors DCT histogram smoothing Anti-forensics Conclusions JPEG image restoration The minimization problem regularization parameter covariance matrix for the k -th kind of patch ���� 64 ���� λ � � ( P i ( x q − x )) t C − 1 x = arg min ˆ { k P i ( x q − x ) 2 x k =1 P i ∈S k � �� � the i -th overlapping patch the k -th group of patch extracting matrices � ���� − log p ( P i x ) } i • The problem is solved using an approximate MAP procedure and a Quantization Constraint Set (QCS) projection JPEG Anti-forensics Using Non-parametric DCT Quantization Noise Estimation and Natural Image Statistics 13 / 24