MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling Matthias Carnein Pascal Schöttle Rainer Böhme 19th November 2015
Introduction uncompressed Q = 15 Q = 100 Block convergence analysis can detect high quality JPEG compression Goal: Extend to color images Image source: Dang-Nguyen et al. (2015)
Outline MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 3 /31 1. Block Convergence 2. Extension To Color 3. Experimental Results 4. Conclusion
Outline MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 3 /31 1. Block Convergence 2. Extension To Color 3. Experimental Results 4. Conclusion
Grayscale JPEG MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 4 /31 Color Conversion Subsampling Block Splitting [ ] yij q ij Discrete Cosine Transformation Quantization Entropy Coding
Block Convergence Lai and Böhme (2013)
Block Convergence stable in t = 2 t = 0 t = 2 t = 1 Lai and Böhme (2013)
MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 6 /31 Ratio of Stable Blocks Number of compressions: 1 Lai and Böhme (2013)
MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 6 /31 Ratio of Stable Blocks Number of compressions: 5 Ratio of stable blocks: r = (#stable #flat)/(#total #flat) Lai and Böhme (2013)
Distribution MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 7 /31 ratio of stable blocks r 100% 80% 60% 40% 20% 0% 0 1 2 3 4 5 6 7 8 9 10 number of compressions t
Outline MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 8 /31 1. Block Convergence 2. Extension To Color 3. Experimental Results 4. Conclusion
Block Convergence MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 9 /31 block is stable if no channel changes block red green blue
MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 10 /31 Ratio of Stable Blocks Number of compressions: 5
MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 10 /31 Ratio of Stable Blocks Number of compressions: 5
MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 10 /31 Ratio of Stable Blocks Number of compressions: 5
Distribution MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 11 /31 ratio of stable blocks r 100% 80% 60% 40% 20% 0% 0 1 2 3 4 5 6 7 8 9 10 number of compressions t
Color JPEG G MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 12 /31 R C r Y C b B Color Conversion Subsampling C b C b C b C b C b Block Splitting Discrete Cosine Transformation Quantization Entropy Coding
MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 13 /31 Simple Sub- and Upsampling Subsampling Upsampling 1/4 1 Independent JPEG Group (2015)
Macro-Blocks MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 14 /31 Observe convergence for area covered by chrominance block
Macro-Blocks MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 14 /31 Observe convergence for area covered by chrominance block
Distribution MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 15 /31 ratio of stable macro-blocks r 100% 80% 60% 40% 20% 0% r (2) r (0) r(1)? r (3) 0 1 2 3 4 5 6 7 8 9 10 number of compressions t
Fancy Upsampling MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 16 /31 Subsampling Upsampling 1/4 1 9/16 3/16 1/16 Independent JPEG Group (2015)
Spill-Over MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 17 /31 fancy upsampling not confined within a macro-block ignore spill-over: stable if block does not change with compression
Spill-Over MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 17 /31 fancy upsampling not confined within a macro-block ignore spill-over: stable if block does not change with compression prevent spill-over: compress each block individually
Distribution MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 18 /31 ratio of stable macro-blocks r 100% 80% 60% 40% 20% 0%? 0 2 4 6 8 10 12 14 16 18 20 number of compressions t
MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 19 /31 Multiple Recompressions Track ratio over multiple recompressions: r 0 r 1... r n Option 1: Machine Learning use as features to train machine learning algorithm relative frequency Option 2: Maximum-Likelihood Empirical Distributions 1 0.5 0 0 0.2 0.4 0.6 0.8 1 ratio of stable blocks r probability density 40 20 Fitted Beta Distributions 0 0 0.2 0.4 0.6 0.8 1 ratio of stable blocks r
Outline MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 20 /31 1. Block Convergence 2. Extension To Color 3. Experimental Results 4. Conclusion
Experimental Setup MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 21 /31 1600 images up to 8 compressions results reported as confusion matrices true detection rate off-by-two ratio Detector output ˆt 0 1 2 3 4 5 6 7 8 9 10 Number of compressions t 0 1 2 3 4 5 6 7 8
Detection Accuracy MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 22 /31 subsampling Random Forest Maximum-likelihood Q rate algorithm DCT true detection off-by-two true detection off-by-two 100 4:4:4 slow 98.61% 10.00% 88.36% 7.64% 100 4:4:4 float 96.35% 9.89% 85.88% 20.45% 100 4:2:0 simple slow 98.75% 4.44% 85.42% 6.10% 100 4:2:0 fancy (ignore) slow 94.12% 15.84% 41.81% 44.84% 100 4:2:0 fancy (prevent) slow 90.06% 23.04% 34.71% 60.46% 100 4:2:0 DCT scaling slow 98.72% 2.17% 93.86% 4.07% 99 4:4:4 slow 98.83% 2.38% 90.21% 6.52% 95 4:4:4 slow 88.62% 32.60% 64.54% 42.11% 90 4:4:4 slow 69.31% 57.15% 29.19% 66.91% 75 4:4:4 slow 49.06% 72.25% 23.67% 79.86% 50 4:4:4 slow 35.69% 82.83% 33.39% 54.86%
Confusion Matrix MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 23 /31 Detector output ˆt Number of compressions t 0 1 2 3 4 5 6 7 8 0 800 1 0 0 0 0 0 0 0 1 0 799 2 0 0 0 0 0 0 2 0 0 798 2 0 0 0 0 0 3 0 0 0 798 2 1 0 0 0 4 0 0 0 0 798 4 1 0 0 5 0 0 0 0 0 795 7 2 0 6 0 0 0 0 0 0 790 14 2 7 0 0 0 0 0 0 2 776 30 8 0 0 0 0 0 0 0 8 746 9 0 0 0 0 0 0 0 0 18 10 0 0 0 0 0 0 0 0 4
Outline MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 24 /31 1. Block Convergence 2. Extension To Color 3. Experimental Results 4. Conclusion
Conclusion MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 25 /31 Block Convergence: use ratio of converged blocks to estimate number of JPEG compressions Extension to Color observe block convergence for macro-blocks adapt to sub- and upsampling algorithm track ratio over multiple recompressions use maximum-likelihood or machine learning Outlook Drawback: assumes repeated compression with the same settings apply to other lossy compression formats
MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling Thank you for your attention! 19th November 2015
References MÜNSTER Forensics of High-Quality JPEG Images with Color Subsampling 27 /31 Dang-Nguyen Duc-Tien Cecilia Pasquini Valentina Conotter and Giulia Boato (2015). RAISE: A Raw Images Dataset for Digital Image Forensics. In: Proceedings of the 6th ACM Multimedia Systems Conference. MMSys 15. Portland Oregon: ACM pp. 219 224. URL: http://mmlab.science.unitn.it/raise/ (visited on 17/11/2015). Independent JPEG Group (2015). libjpeg. URL: http://www.ijg.org/ (visited on 17/11/2015). Lai ShiYue and Rainer Böhme (2013). Block convergence in repeated transform coding: JPEG-100 forensics carbon dating and tamper detection. In: IEEE International Conference on Acoustics Speech and Signal Processing (ICASSP) pp. 3028 3032.