Towards Digital Image Anti-Forensicsvia Image Restoration
A Ph.D. thesis defense
by
Wei FAN
supervised by:Kai WANG, Francois CAYRE, and Jean-Marc BROSSIER at GIPSA-lab,
and Zhang XIONG at Beihang University
30/04/2015
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Seeing is believing.A picture is worth a thousand words.
Manipulation
Dutch painter Johannes Vermeer,around 1665 Worth1000 user bigchopper, 2014
Wei FAN 2 / 50I Original painting information: http://en.wikipedia.org/wiki/Girl_with_a_Pearl_EarringI Image forgery source: http://www.worth1000.com/entries/740270/girl-with-the-beats
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Image Forensics:Restore Some Trust to Digital Images
Image ForensicsTo analyze a given digital image so as to detect whether it is aforgery, to identify its origin, to trace its processing history, or toreveal latent details invisible to human naked eyes.
Forensics{
Active Forensics (Fragile Watermarking)Passive Forensics, often directly referred to as Forensics
Wei FAN 3 / 50I Fourandsix Technologies, Inc. http://www.fourandsix.com/about-us/I V. Conotter. “Active and passive multimedia forensics”. PhD thesis. University of Trento, 2011
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Every coin has two sides.
cryptography vs. cryptanalysissteganography vs. steganalysis
Image Anti-ForensicsTo expose the limitations of forensic methods, with the ultimategoal to develop more trustworthy forensics.
Wei FAN 4 / 50I R. Bohme and M. Kirchner. “Counter-forensics: attacking image forensics”. Digital Image Forensics, Springer,
2013, pp. 327-366
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Research Context
Original image
Wei FAN 5 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Research Context
Original image Image forgery
Manipulation
Wei FAN 5 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Research Context
Original image Image forgery
Manipulation
Forensic tool
Original image Image forgery
Manipulation
Forensic tool
Wei FAN 5 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Research Context
Original image Image forgery
Manipulation
Forensic tool
forensic feature differs
Original image Image forgery
Manipulation
Forensic tool
forensic feature differs
Wei FAN 5 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Research Context
Original image Image forgery
Manipulation
Forensic tool
forensic feature differs
Image Forgery Detected! (Correctly Classified)
Original image Image forgery
Manipulation
Forensic tool
forensic feature differs
Image Forgery Detected! (Correctly Classified)
Wei FAN 5 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Research Context
Original image Image forgery
Manipulation
Anti-forensic image
Anti-forensics
Wei FAN 5 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Research Context
Original image Image forgery
Manipulation
Anti-forensic image
Anti-forensics
Forensic tool
Original image Image forgery
Manipulation
Anti-forensic image
Anti-forensics
Forensic tool
Wei FAN 5 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Research Context
Original image Image forgery
Manipulation
Anti-forensic image
Anti-forensics
Forensic toolforensic feature resembles
Original image Image forgery
Manipulation
Anti-forensic image
Anti-forensics
Forensic toolforensic feature resembles
Wei FAN 5 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Research Context
Original image Image forgery
Manipulation
Anti-forensic image
Anti-forensics
Forensic toolforensic feature resembles
Original Image Detected! (Wrongly Classified)
Original image Image forgery
Manipulation
Anti-forensic image
Anti-forensics
Forensic toolforensic feature resembles
Original Image Detected! (Wrongly Classified)
Wei FAN 5 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Research Context
Original image Image forgery
Manipulation
Anti-forensic image
Anti-forensics
Thesis objective: to design image anti-forensic methods.
Wei FAN 5 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Research Context
Original image Image forgery
Manipulation
Anti-forensic image
Anti-forensics
Thesis objective: to design image anti-forensic methods.
Why? the development of trustworthy image forensics.
Wei FAN 5 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Farid’s Classification of Image ForensicsImage Forensics
Format Pixel Camera Physically Geometric
JPEG
SPHIT
· · ·
Medianfiltering
Resampling
Splicing
· · ·
Chromaticaberration
CFA
PRNU
· · ·
Lightingdirection
Lightingenviron-
ment
Illuminationcolor
· · ·
Principalpoint
Shadows
Reflections
· · ·
Wei FAN 6 / 50
I H. Farid, “A Survey of Image Forgery Detection,” IEEE Signal Processing Magazine, 2009
I http://w3techs.com/technologies/overview/image_format/allI M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065I M. Kirchner and R. Rohme. “Hiding traces of resampling in digital images”. IEEE TIFS 3, 4 (2008), pp. 582-592
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Farid’s Classification of Image ForensicsImage Forensics
Format Pixel Camera Physically Geometric
JPEG
SPHIT
· · ·
Medianfiltering
Resampling
Splicing
· · ·
Chromaticaberration
CFA
PRNU
· · ·
Lightingdirection
Lightingenviron-
ment
Illuminationcolor
· · ·
Principalpoint
Shadows
Reflections
· · ·
Wei FAN 6 / 50
I H. Farid, “A Survey of Image Forgery Detection,” IEEE Signal Processing Magazine, 2009
I http://w3techs.com/technologies/overview/image_format/allI M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065I M. Kirchner and R. Rohme. “Hiding traces of resampling in digital images”. IEEE TIFS 3, 4 (2008), pp. 582-592
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Farid’s Classification of Image ForensicsImage Forensics
Format Pixel Camera Physically Geometric
JPEG
SPHIT
· · ·
Medianfiltering
Resampling
Splicing
· · ·
Chromaticaberration
CFA
PRNU
· · ·
Lightingdirection
Lightingenviron-
ment
Illuminationcolor
· · ·
Principalpoint
Shadows
Reflections
· · ·
JPEG is the most widely used image format on Internet.
Wei FAN 6 / 50
I H. Farid, “A Survey of Image Forgery Detection,” IEEE Signal Processing Magazine, 2009
I http://w3techs.com/technologies/overview/image_format/all
I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.1050-1065
I M. Kirchner and R. Rohme. “Hiding traces of resampling in digital images”. IEEE TIFS 3, 4 (2008), pp. 582-592
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Farid’s Classification of Image ForensicsImage Forensics
Format Pixel Camera Physically Geometric
JPEG
SPHIT
· · ·
Medianfiltering
Resampling
Splicing
· · ·
Chromaticaberration
CFA
PRNU
· · ·
Lightingdirection
Lightingenviron-
ment
Illuminationcolor
· · ·
Principalpoint
Shadows
Reflections
· · ·
JPEG is the most widely used image format on Internet.
Median filtering is a widely used image processing operationfor, e.g., denoising, smoothing, etc.
Median filtering is also used for anti-forensic purposes, e.g.,JPEG image deblocking, disguising resampling artifacts.
Wei FAN 6 / 50
I H. Farid, “A Survey of Image Forgery Detection,” IEEE Signal Processing Magazine, 2009
I http://w3techs.com/technologies/overview/image_format/allI M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065I M. Kirchner and R. Rohme. “Hiding traces of resampling in digital images”. IEEE TIFS 3, 4 (2008), pp. 582-592
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Creating A Composite JPEG ImageJPEG image with q1 JPEG image with q2
resulting double JPEG compressed image
image composition, JPEG compression again with q3
JPEG compressedtwice with q1 and q3
JPEG compressedtwice with q2 and q3
Wei FAN 7 / 50I T. Bianchi and A. Piva. “Image forgery localization via block-grained analysis of JPEG artifacts”. IEEE TIFS 7,
3 (2012), pp. 1003-1017
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Research Context
Creating A Composite JPEG ImageJPEG image with q1 JPEG image with q2
resulting double JPEG compressed image
image composition, JPEG compression again with q3
JPEG compressedtwice with q1 and q3
JPEG compressedtwice with q2 and q3
−60
−40
−20
0
20
likelihood map
Wei FAN 7 / 50I T. Bianchi and A. Piva. “Image forgery localization via block-grained analysis of JPEG artifacts”. IEEE TIFS 7,
3 (2012), pp. 1003-1017
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Methodology
Analysis of Current Image Anti-Forensics
1 JPEG anti-forensicsDithering (≈ noise addition) for DCT histogram smoothingMedian filtering for JPEG deblocking
2 Median filtering anti-forensicsSharpening filteringDithering (≈ noise addition)Noise injection
Wei FAN 8 / 50
I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.1050-1065
I M. Fontani and M. Barni. “Hiding traces of median filtering in digital images”. In: Proc. EUSIPCO: IEEE, 2012,pp. 1239-1243
I Z.-H. Wu, et al. “Anti-forensics of median filtering”. In: Proc. ICASSP. 2013, pp. 3043-3047I D. T. Dang-Nguyen et al. “Counter-forensics of median filtering”. In: Proc. MMSP. 2013, pp. 260-265
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Methodology
Analysis of Current Image Anti-Forensics
1 JPEG anti-forensicsDithering (≈ noise addition) for DCT histogram smoothingMedian filtering for JPEG deblocking
2 Median filtering anti-forensicsSharpening filteringDithering (≈ noise addition)Noise injection
Wei FAN 8 / 50
I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.1050-1065
I M. Fontani and M. Barni. “Hiding traces of median filtering in digital images”. In: Proc. EUSIPCO: IEEE, 2012,pp. 1239-1243
I Z.-H. Wu, et al. “Anti-forensics of median filtering”. In: Proc. ICASSP. 2013, pp. 3043-3047I D. T. Dang-Nguyen et al. “Counter-forensics of median filtering”. In: Proc. MMSP. 2013, pp. 260-265
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Methodology
Analysis of Current Image Anti-Forensics
1 JPEG anti-forensicsDithering (≈ noise addition) for DCT histogram smoothingMedian filtering for JPEG deblocking
2 Median filtering anti-forensicsSharpening filteringDithering (≈ noise addition)Noise injection
Current anti-forensic methods mainly use simple imageprocessing, e.g., filtering, noise addition, etc.
1 Image quality is a concern2 Can be detected by advanced forensic algorithms
Wei FAN 8 / 50
I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.1050-1065
I M. Fontani and M. Barni. “Hiding traces of median filtering in digital images”. In: Proc. EUSIPCO: IEEE, 2012,pp. 1239-1243
I Z.-H. Wu, et al. “Anti-forensics of median filtering”. In: Proc. ICASSP. 2013, pp. 3043-3047I D. T. Dang-Nguyen et al. “Counter-forensics of median filtering”. In: Proc. MMSP. 2013, pp. 260-265
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Methodology
Leveraging on Image Restoration
Image RestorationTo estimate the “clean” original image from the corrupted image,usually via solving an ill-posed inverse problem.
Image Anti-Forensics vs. Image RestorationSimilarities
Process the degraded image to approximate the original oneRequire high quality of the processed image
DifferencesAnti-forensics: good forensic undetectability is a mustRestoration: high image quality is the goal
Wei FAN 9 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Methodology
Leveraging on Image Restoration
Image RestorationTo estimate the “clean” original image from the corrupted image,usually via solving an ill-posed inverse problem.
Image Anti-Forensics vs. Image RestorationSimilarities
Process the degraded image to approximate the original oneRequire high quality of the processed image
DifferencesAnti-forensics: good forensic undetectability is a mustRestoration: high image quality is the goal
Proposed Methodology:
Employ MAP estimation (or one of its variants)Adopt & enrich statistical models from image restorationIntegrate some anti-forensic terms/strategies
Wei FAN 9 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Preliminaries
Natural Image Datasets
1 JPEG compressionUCID: 1338 original non-compressed images with size 512×384
2 Median filteringMFTE, MFTR, MFPE: 1607 original, never resampled, non-compressed images with size 512× 512
Wei FAN 10 / 50
I G. Schaefer and M. Stich. “UCID - an uncompressed colour image database”. In: Proc. SPIE, 2004, pp. 472-480I ftp://firewall.teleco.uvigo.es:27244/DS_01_UTFI.zipI ftp://lesc.dinfo.unifi.it/pub/Public/JPEGloc/dataset/
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Preliminaries
Evaluation MetricsAnti-Forensics Objective: Good Undetectability & High Image Quality
1 Forensic UndetectabilityArea Under Curve (AUC)
2 Image QualityPeak Signal-to-Noise Ratio (PSNR)Structural SIMilarity (SSIM)
3 Histogram RecoveryKullback-Leibler divergence (KL divergence)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
False positive rate
Truepositive
rate
Random guess
Goal:
ROC curve → random guess line AUC → 0.5
Goal:
the higher the PSNR/SSIM value is, the better(reference: the original image)
Goal:
KL divergence → 0(reference: certain histogram constructed from the original image)
original image processed image
JPEG compression/ median filtering(+ anti-forensics)
with a replacement rate
composite image forgery
Wei FAN 11 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Preliminaries
Evaluation MetricsAnti-Forensics Objective: Good Undetectability & High Image Quality
1 Forensic UndetectabilityArea Under Curve (AUC)
2 Image QualityPeak Signal-to-Noise Ratio (PSNR)Structural SIMilarity (SSIM)
3 Histogram RecoveryKullback-Leibler divergence (KL divergence)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
False positive rate
Truepositive
rate
Random guess
Goal:
ROC curve → random guess line AUC → 0.5
Goal:
the higher the PSNR/SSIM value is, the better(reference: the original image)
Goal:
KL divergence → 0(reference: certain histogram constructed from the original image)
original image processed image
JPEG compression/ median filtering(+ anti-forensics)
with a replacement rate
composite image forgery
Wei FAN 11 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Preliminaries
Evaluation MetricsAnti-Forensics Objective: Good Undetectability & High Image Quality
1 Forensic UndetectabilityArea Under Curve (AUC)
2 Image QualityPeak Signal-to-Noise Ratio (PSNR)Structural SIMilarity (SSIM)
3 Histogram RecoveryKullback-Leibler divergence (KL divergence)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
False positive rate
Truepositive
rate
Random guess
Goal:
ROC curve → random guess line AUC → 0.5
Goal:
the higher the PSNR/SSIM value is, the better(reference: the original image)
Goal:
KL divergence → 0(reference: certain histogram constructed from the original image)
original image processed image
JPEG compression/ median filtering(+ anti-forensics)
with a replacement rate
composite image forgery
Wei FAN 11 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Preliminaries
Evaluation MetricsAnti-Forensics Objective: Good Undetectability & High Image Quality
1 Forensic UndetectabilityArea Under Curve (AUC)
2 Image QualityPeak Signal-to-Noise Ratio (PSNR)Structural SIMilarity (SSIM)
3 Histogram RecoveryKullback-Leibler divergence (KL divergence)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
False positive rate
Truepositive
rate
Random guess
Goal:
ROC curve → random guess line AUC → 0.5
Goal:
the higher the PSNR/SSIM value is, the better(reference: the original image)
Goal:
KL divergence → 0(reference: certain histogram constructed from the original image)
original image processed image
JPEG compression/ median filtering(+ anti-forensics)
with a replacement rate
composite image forgery
Wei FAN 11 / 50I Z. Wang, et al. “Image quality assessment: from error visibility to structural similarity”. IEEE TIP 13, 4 (2004),
pp. 600-612
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Preliminaries
Evaluation MetricsAnti-Forensics Objective: Good Undetectability & High Image Quality
1 Forensic UndetectabilityArea Under Curve (AUC)
2 Image QualityPeak Signal-to-Noise Ratio (PSNR)Structural SIMilarity (SSIM)
3 Histogram RecoveryKullback-Leibler divergence (KL divergence)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
False positive rate
Truepositive
rate
Random guess
Goal:
ROC curve → random guess line AUC → 0.5
Goal:
the higher the PSNR/SSIM value is, the better(reference: the original image)
Goal:
KL divergence → 0(reference: certain histogram constructed from the original image)
original image processed image
JPEG compression/ median filtering(+ anti-forensics)
with a replacement rate
composite image forgery
Wei FAN 11 / 50I S. Kullback and R. A. Leibler. “On information and sufficiency”. Annals of Mathematical Statistics 22, 1 (1951),
pp. 49-86
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
1 Introduction
2 JPEG Anti-ForensicsTV-based JPEG deblockingPerceptual DCT histogram smoothingUsing a sophisticated image model
3 Median Filtering Anti-Forensics
4 Conclusions & Perspectives
Wei FAN 12 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
JPEG Artifacts
Blocking artifacts
DCT coefficient value
DCT
coeffi
cientfrequen
cy
−300 −100 100 3000
0.02
0.04
0.06
DCT coefficient value
DCT
coeffi
cientfrequen
cy
−300 −100 100 3000
0.2
0.4
0.6
0.8
Comb-like quantization artifactsWei FAN 13 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
1. TV-Based JPEG Deblocking
Image Total VariationPixel classification:
A: shaded pixels
B: the other pixels
For original image:variation (A) ≈ variation (B)
Total Variation (TV)Simple, but effective image modelWidely used in image denoising, JPEG post-processing, etc.Here, to measure “blocking”, by pixel classification
Wei FAN 14 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
1. TV-Based JPEG Deblocking
TV-Based JPEG Anti-ForensicsDeblocking
arg minU∈ S︸︷︷︸
constraint image space
TV︷ ︸︸ ︷∑
1≤i≤H ,1≤j≤W υi,j +α
∣∣∣∣∣∑Ui,j∈A υi,j −∑
Ui,j∈B υi,j
∣∣∣∣∣︸ ︷︷ ︸TV-based blocking measurement
TV, image prior
TV-based blocking measurement, deblocking
De-Calibrationarg min
U
28∑k=1
∣∣∣var(DkU)− var(DkUcal)∣∣∣
Ucal : crop the first 4 pixels of U both horizontally and vertically
Wei FAN 15 / 50
I F. Alter, et al. “Adapted total variation for artifact free decompression of JPEG images”. JMIV 23, 2 (2005),pp. 199-211
I S. Lai and R. Bohme, “Countering counter-forensics: the case of JPEG compression,” In: Proc. IH, 2011, pp.285-298
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
1. TV-Based JPEG Deblocking
TV-Based JPEG Anti-ForensicsDeblocking
arg minU∈ S︸︷︷︸
constraint image space
TV︷ ︸︸ ︷∑
1≤i≤H ,1≤j≤W υi,j +α
∣∣∣∣∣∑Ui,j∈A υi,j −∑
Ui,j∈B υi,j
∣∣∣∣∣︸ ︷︷ ︸TV-based blocking measurement
TV, image prior
TV-based blocking measurement, deblocking
De-Calibrationarg min
U
28∑k=1
∣∣∣var(DkU)− var(DkUcal)∣∣∣
Ucal : crop the first 4 pixels of U both horizontally and vertically
Wei FAN 15 / 50
I F. Alter, et al. “Adapted total variation for artifact free decompression of JPEG images”. JMIV 23, 2 (2005),pp. 199-211
I S. Lai and R. Bohme, “Countering counter-forensics: the case of JPEG compression,” In: Proc. IH, 2011, pp.285-298
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
1. TV-Based JPEG Deblocking
TV-Based JPEG Anti-ForensicsDeblocking
arg minU∈ S︸︷︷︸
constraint image space
TV︷ ︸︸ ︷∑
1≤i≤H ,1≤j≤W υi,j +α
∣∣∣∣∣∑Ui,j∈A υi,j −∑
Ui,j∈B υi,j
∣∣∣∣∣︸ ︷︷ ︸TV-based blocking measurement
TV, image prior
TV-based blocking measurement, deblocking
De-Calibrationarg min
U
28∑k=1
∣∣∣var(DkU)− var(DkUcal)∣∣∣
Ucal : crop the first 4 pixels of U both horizontally and vertically
Wei FAN 15 / 50
I F. Alter, et al. “Adapted total variation for artifact free decompression of JPEG images”. JMIV 23, 2 (2005),pp. 199-211
I S. Lai and R. Bohme, “Countering counter-forensics: the case of JPEG compression,” In: Proc. IH, 2011, pp.285-298
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
1. TV-Based JPEG Deblocking
TV-Based JPEG Anti-ForensicsDeblocking
arg minU∈ S︸︷︷︸
constraint image space
TV︷ ︸︸ ︷∑
1≤i≤H ,1≤j≤W υi,j +α
∣∣∣∣∣∑Ui,j∈A υi,j −∑
Ui,j∈B υi,j
∣∣∣∣∣︸ ︷︷ ︸TV-based blocking measurement
TV, image prior
TV-based blocking measurement, deblocking
De-Calibrationarg min
U
28∑k=1
∣∣∣var(DkU)− var(DkUcal)∣∣∣
Ucal : crop the first 4 pixels of U both horizontally and vertically
Wei FAN 15 / 50
I F. Alter, et al. “Adapted total variation for artifact free decompression of JPEG images”. JMIV 23, 2 (2005),pp. 199-211
I S. Lai and R. Bohme, “Countering counter-forensics: the case of JPEG compression,” In: Proc. IH, 2011, pp.285-298
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
1. TV-Based JPEG Deblocking
Experimental Results
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KF
KLuo
KQLuo
KV
KL
K1U
K2U
Random guess
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KF
KLuo
KQLuo
KV
KL
K1U
K2U
Random guess
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KF
KLuo
KQLuo
KV
KL
K1U
K2U
Random guess
J , JPEG FJSqSb
, state-of-the-art FJ0 , proposed
J FJSqSb
FJ0
PSNR [dB] 37.0999 30.4591 35.4814SSIM 0.9919 0.9509 0.9843
SummaryFJ
0 has a PSNR gain of 5 dBover FJ
SqSb
The quality of FJ0 is slightly less
good than JROC curves achieved by FJ
0 arethe closest to the random guess
Wei FAN 16 / 50I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
1. TV-Based JPEG Deblocking
Experimental Results
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KF
KLuo
KQLuo
KV
KL
K1U
K2U
Random guess
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KF
KLuo
KQLuo
KV
KL
K1U
K2U
Random guess
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KF
KLuo
KQLuo
KV
KL
K1U
K2U
Random guess
J , JPEG FJSqSb
, state-of-the-art FJ0 , proposed
J FJSqSb
FJ0
PSNR [dB] 37.0999 30.4591 35.4814SSIM 0.9919 0.9509 0.9843
SummaryFJ
0 has a PSNR gain of 5 dBover FJ
SqSb
The quality of FJ0 is slightly less
good than JROC curves achieved by FJ
0 arethe closest to the random guess
Wei FAN 16 / 50I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
1. TV-Based JPEG Deblocking
Example Results
— PSNR = 30.2841 dB PSNR = 26.4496 dB PSNR = 29.8084 dB
DCT coefficient value
DCT
coeffi
cientfrequency
−300 −200 −100 0 100 200 3000
0.01
0.02
0.03
0.04
DCT coefficient value
DCT
coeffi
cientfrequen
cy
−300 −200 −100 0 100 200 3000
0.1
0.2
0.3
DCT coefficient value
DCT
coeffi
cientfrequency
−300 −200 −100 0 100 200 3000
0.01
0.02
0.03
DCT coefficient value
DCT
coeffi
cientfrequency
−300 −200 −100 0 100 200 3000
0.02
0.04
0.06
I, original J , JPEG FJSqSb , state-of-the-art FJ
0 , proposed
Wei FAN 17 / 50I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
1. TV-Based JPEG Deblocking
Example Results
— PSNR = 30.2841 dB PSNR = 26.4496 dB PSNR = 29.8084 dB
DCT coefficient value
DCT
coeffi
cientfrequency
−300 −200 −100 0 100 200 3000
0.01
0.02
0.03
0.04
DCT coefficient value
DCT
coeffi
cientfrequen
cy
−300 −200 −100 0 100 200 3000
0.1
0.2
0.3
DCT coefficient value
DCT
coeffi
cientfrequency
−300 −200 −100 0 100 200 3000
0.01
0.02
0.03
DCT coefficient value
DCT
coeffi
cientfrequency
−300 −200 −100 0 100 200 3000
0.02
0.04
0.06
I, original J , JPEG FJSqSb , state-of-the-art FJ
0 , proposed
Wei FAN 17 / 50I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
1. TV-Based JPEG Deblocking
Potential Weakness in DCT Histogram
DCT coefficient value
DCTcoeffi
cientfrequency
−100 −50 0 50 1000
0.02
0.04
0.06
0.08
FJSqSb , state-of-the-art
DCT coefficient value
DCTcoeffi
cientfrequency
−150 −100 −50 0 50 100 1500
0.04
0.08
0.12
FJ0 , proposed
Usually happens for FJ0 in the mid-frequency subbands.
Wei FAN 18 / 50I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Improved JPEG Anti-Forensics withPerceptual DCT Histogram Smoothing
DCT coefficient value
DCTcoeffi
cientfrequency
−150 −100 −50 0 50 100 1500
0.04
0.08
0.12
FJ0
J first-roundTV-based deblocking
perceptual DCThistogram smoothing
second-roundTV-based deblockingde-calibrationFJ
FJb
FJbq
FJbqb
Handle JPEG artifacts separately in spatial and DCT domainsSlightly different parameter settings for TV-based deblocking
Wei FAN 19 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Disadvantages of Global Laplacian inModeling DCT Coefficient
DCT coefficient value
DCT
coeffi
cientfrequen
cy
−150 −100 −50 0 50 100 1500
0.04
0.08
0.12
0.16
Kurtosis of the Laplacian distribution: 6
Average kurtosis of AC components ofUCID images: 19.99� 6
93.68% of AC components have kurtosisvalue higher than 6
Robertson and Stevenson’s ConclusionQuantization bin 0: Laplacian distributionOther bins: uniform distribution
Wei FAN 20 / 50I M. A. Robertson and R. L. Stevenson. “DCT quantization noise in compressed images”. IEEE TCSVT 15, 1
(2005), pp. 27-38
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
After First-Round TV-Based Deblocking
DCT coefficient value
DCT
coeffi
cientfrequen
cy
−150 −100 −50 0 50 100 1500
0.2
0.4
0.6
J , JPEGDCT coefficient value
DCT
coeffi
cientfrequen
cy
−150 −100 −50 0 50 100 1500
0.04
0.08
0.12
0.16
FJb , deblocked
Comb-like quantization artifacts are partly removedBut gaps still remain to some extent
Wei FAN 21 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Adaptive Local Laplacian Model
1 Find the parameter λb of the Laplacian distribution for eachquantization bin
2 If not possible, use the uniform distribution
Finding λb
λb = arg minλ−b ≤λ≤λ
+b
B+r,cQr,c+
⌊Qr,c
2
⌋∑
k=B−r,cQr,c−⌊
Qr,c2
⌋wk ×(H X
r ,c(k)− P(Y = k))2
Wei FAN 22 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Perceptual DCT Histogram Mapping
DCT coefficient value
DCT
coeffi
cientfrequen
cy
−150 −100 −50 0 50 100 1500
0.02
0.04
0.06
0.08
0.1
Distribution target
Minimize the total SSIM value loss
Assignment Problem (for each quantization bin b)
∑o∈Ob
weight function W : Ob × Tb → R︷︸︸︷W (o, f︸︷︷︸bijection f : Ob → Tb
(o))
Ob: the to-be-modified DCT coefficientsof FJ
b
T b: target DCT coefficients values inthe dithered histogram
Wei FAN 23 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Example Results
J , SSIM = 0.9809 FJV , SSIM = 0.9509
FJSq , SSIM = 0.9610 FJ
bq, SSIM = 0.9731
Summary
Less noise present in FJbq
Especially advantageousfor FJ
bq in the relativelysmooth image areas
I G. Valenzise, et al. “The cost of JPEGcompression anti-forensics”. In: Proc.ICASSP. 2011, pp. 1884-1887
I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”.IEEE TIFS 6, 3 (2011), pp. 1050-1065
Wei FAN 24 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Example Results
J , SSIM = 0.9809 FJV , SSIM = 0.9509
FJSq , SSIM = 0.9610 FJ
bq, SSIM = 0.9731
Summary
Less noise present in FJbq
Especially advantageousfor FJ
bq in the relativelysmooth image areas
I G. Valenzise, et al. “The cost of JPEGcompression anti-forensics”. In: Proc.ICASSP. 2011, pp. 1884-1887
I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”.IEEE TIFS 6, 3 (2011), pp. 1050-1065
Wei FAN 24 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Histogram Recovery Comparison — FJSq vs. FJ
bq
Difference of KL divergence values:
r c 1 2 3 4 5 6 7 81 0.0001 0.0065 0.0118 0.0278 0.0493 0.0634 0.0663 0.06562 0.0042 0.0166 0.0229 0.0363 0.0447 0.0565 0.0504 0.03693 0.0161 0.0208 0.0291 0.0442 0.0573 0.0665 0.0634 0.04974 0.0200 0.0317 0.0409 0.0470 0.0658 0.0802 0.0553 0.04465 0.0357 0.0395 0.0522 0.0678 0.0764 0.0930 0.0927 0.08566 0.0441 0.0383 0.0642 0.0610 0.0726 0.0769 0.0806 0.09577 0.0538 0.0442 0.0678 0.0595 0.0879 0.0809 0.0948 0.09758 0.0619 0.0545 0.0697 0.0528 0.0927 0.0880 0.0854 0.0722
Average: 0.0552, standard deviation: 0.0249
FJbq constantly outperforms FJ
Sq
Wei FAN 25 / 50I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Histogram Recovery Comparison — FJSq vs. FJ
bq
Difference of KL divergence values:
r c 1 2 3 4 5 6 7 81 0.0001 0.0065 0.0118 0.0278 0.0493 0.0634 0.0663 0.06562 0.0042 0.0166 0.0229 0.0363 0.0447 0.0565 0.0504 0.03693 0.0161 0.0208 0.0291 0.0442 0.0573 0.0665 0.0634 0.04974 0.0200 0.0317 0.0409 0.0470 0.0658 0.0802 0.0553 0.04465 0.0357 0.0395 0.0522 0.0678 0.0764 0.0930 0.0927 0.08566 0.0441 0.0383 0.0642 0.0610 0.0726 0.0769 0.0806 0.09577 0.0538 0.0442 0.0678 0.0595 0.0879 0.0809 0.0948 0.09758 0.0619 0.0545 0.0697 0.0528 0.0927 0.0880 0.0854 0.0722
all positive
Average: 0.0552, standard deviation: 0.0249
FJbq constantly outperforms FJ
Sq
Wei FAN 25 / 50I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Forensic Undetectability
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KF
KLuo
KQLuo
KV
KL
K1U
K2U
Random guess
FJ
Replacement rate
AUC
0 0.1 0.2 0.3 0.4 0.5 0.6 0.8 10.5
0.6
0.7
0.8
0.9
1
J
FJSq
FJSqSb
FJV
FJSu
FJ0
FJ
KS100Li
Replacement rate
AUC
0 0.1 0.2 0.3 0.4 0.5 0.6 0.8 10.5
0.6
0.7
0.8
0.9
1
J
FJSq
FJSqSb
FJV
FJSu
FJ0
FJ
KS686P
Scalar-based detectors:ROC curves achieved by FJ are close to the random guess
SVM-based detectors:At low image replacement rate: good undetectabilityFJ outperforms the state-of-the-art anti-forensic JPEG images
Wei FAN 26 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Forensic Undetectability
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KF
KLuo
KQLuo
KV
KL
K1U
K2U
Random guess
FJ
Replacement rate
AUC
0 0.1 0.2 0.3 0.4 0.5 0.6 0.8 10.5
0.6
0.7
0.8
0.9
1
J
FJSq
FJSqSb
FJV
FJSu
FJ0
FJ
KS100Li
Replacement rate
AUC
0 0.1 0.2 0.3 0.4 0.5 0.6 0.8 10.5
0.6
0.7
0.8
0.9
1
J
FJSq
FJSqSb
FJV
FJSu
FJ0
FJ
KS686P
Scalar-based detectors:ROC curves achieved by FJ are close to the random guess
SVM-based detectors:At low image replacement rate: good undetectabilityFJ outperforms the state-of-the-art anti-forensic JPEG images
FJ FJ
Wei FAN 26 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Example Results
I, — J , PSNR = 31.8094 dB FJSq Sb
, PSNR = 26.9808 dB FJ , PSNR = 31.5878 dB
DCT coefficient value
DCT
coeffi
cientfrequen
cy
−400 −300 −200 −100 0 100 200 300 4000
0.01
0.02
0.03
0.04
DCT coefficient value
DCT
coeffi
cien
tfreq
uen
cy
−80 −60 −40 −20 0 20 40 60 800
0.05
0.1
0.15
0.2
0.25
DCT coefficient value
DCT
coeffi
cien
tfreq
uen
cy
−60 −40 −20 0 20 40 600
0.05
0.1
0.15
0.2
0.25
0.3
0.35
DCT coefficient value
DCT
coeffi
cien
tfreq
uen
cy
−15 −10 −5 0 5 10 150
0.05
0.1
0.15
0.2
0.25
0.3
0.35
(2, 2) (6, 4) (3, 7) (8, 8)
Wei FAN 27 / 50I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Example Results
I, — J , PSNR = 31.8094 dB FJSq Sb
, PSNR = 26.9808 dB FJ , PSNR = 31.5878 dB
DCT coefficient value
DCT
coeffi
cientfrequen
cy
−400 −300 −200 −100 0 100 200 300 4000
0.01
0.02
0.03
0.04
DCT coefficient value
DCT
coeffi
cien
tfreq
uen
cy
−80 −60 −40 −20 0 20 40 60 800
0.05
0.1
0.15
0.2
0.25
DCT coefficient value
DCT
coeffi
cien
tfreq
uen
cy
−60 −40 −20 0 20 40 600
0.05
0.1
0.15
0.2
0.25
0.3
0.35
DCT coefficient value
DCT
coeffi
cien
tfreq
uen
cy
−15 −10 −5 0 5 10 150
0.05
0.1
0.15
0.2
0.25
0.3
0.35
(2, 2) (6, 4) (3, 7) (8, 8)
Wei FAN 27 / 50I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Application: Disguising Non-Aligned DoubleJPEG Compression
QF2
AUC
50 60 70 80 90 1000.5
0.6
0.7
0.8
0.9
1NA-DJPG-TNA-DJPG-S
NA-DJPG-SS
NA-DJPG-V
NA-DJPG-SuNA-DJPG-F0
NA-DJPG-F
-T -S -SS -V -Su -F0 -FPSNR [dB] 34.6098 32.7958 30.2825 32.4824 30.9379 33.8345 34.0929
SSIM 0.9319 0.8650 0.8487 0.8864 0.8898 0.9168 0.9222
without perceptual DCThistogram smoothing
with perceptual DCThistogram smoothing
Wei FAN 28 / 50I T. Bianchi and A. Piva. “Detection of nonaligned double JPEG compression based on integer periodicity maps”.
IEEE TIFS 7, 2 (2012), pp. 842-848
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
2. Perceptual DCT Histogram Smoothing
Creating A Composite JPEG ImageJPEG image with q1 JPEG image with q2
resulting double JPEG compressed image
image composition, JPEG compression again with q3
JPEG compressedtwice with q1 and q3
JPEG compressedtwice with q2 and q3
−60
−40
−20
0
20
without anti-forensics
−30
−20
−10
0
10
with anti-forensics
Wei FAN 29 / 50I T. Bianchi and A. Piva. “Image forgery localization via block-grained analysis of JPEG artifacts”. IEEE TIFS 7,
3 (2012), pp. 1003-1017
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
JPEG Image Quality Enhancement
AssumptionsJPEG compression: y = x + nq
nq : a random quantity, 0-mean multivariate Gaussiannq and x are independent
MAP criterionx = arg max
xp(x|y) = arg max
xp(y|x)p(x) = arg max
xp(nq)p(x)
Wei FAN 30 / 50I M. A. Robertson and R. L. Stevenson. “DCT quantization noise in compressed images”. IEEE TCSVT 15, 1
(2005), pp. 27-38
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
Proposed Optimization Problem
ModelsImage prior: GMM (Gaussian Mixture Model)
Framework: Expected Patch Log Likelihood (EPLL)
Cost Function
arg minu
{
regularization parameter︷︸︸︷λ
2
64∑k=1
∑Pi∈Sk︸ ︷︷ ︸
the k-th group of patch extracting matrices
(Pi(y− u))t
covariance matrix for the k-th kind of patch︷︸︸︷C−1
k Pi(y− u)
−∑
ilog p(the i-th overlapping patch︷︸︸︷
Piu)}
Wei FAN 31 / 50I D. Zoran and Y. Weiss. “From learning models of natural image patches to whole image restoration”. In: Proc.
ICCV. 2011, pp. 479-486
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
Proposed Optimization Problem
ModelsImage prior: GMM (Gaussian Mixture Model)
Framework: Expected Patch Log Likelihood (EPLL)
Cost Function
arg minu
{
regularization parameter︷︸︸︷λ
2
64∑k=1
∑Pi∈Sk︸ ︷︷ ︸
the k-th group of patch extracting matrices
(Pi(y− u))t
covariance matrix for the k-th kind of patch︷︸︸︷C−1
k Pi(y− u)
−∑
ilog p(the i-th overlapping patch︷︸︸︷
Piu)}
JPEG compression model
Wei FAN 31 / 50I D. Zoran and Y. Weiss. “From learning models of natural image patches to whole image restoration”. In: Proc.
ICCV. 2011, pp. 479-486
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
Proposed Optimization Problem
ModelsImage prior: GMM (Gaussian Mixture Model)
Framework: Expected Patch Log Likelihood (EPLL)
Cost Function
arg minu
{
regularization parameter︷︸︸︷λ
2
64∑k=1
∑Pi∈Sk︸ ︷︷ ︸
the k-th group of patch extracting matrices
(Pi(y− u))t
covariance matrix for the k-th kind of patch︷︸︸︷C−1
k Pi(y− u)
−∑
ilog p(the i-th overlapping patch︷︸︸︷
Piu)}
JPEG compression model
Image prior using GMM under EPLL
Wei FAN 31 / 50I D. Zoran and Y. Weiss. “From learning models of natural image patches to whole image restoration”. In: Proc.
ICCV. 2011, pp. 479-486
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
Experimental Results (PSNR [dB])
JPEG image FoE-based Proposed
LenaQ1 30.71 31.95 32.06Q2 30.08 31.44 31.48Q3 27.45 28.83 28.94
PeppersQ1 30.72 32.04 32.09Q2 30.17 31.61 31.59Q3 27.66 29.35 29.40
BarbaraQ1 25.95 26.65 26.94Q2 25.60 26.31 26.56Q3 24.05 24.86 25.00
BaboonQ1 24.32 24.77 24.84Q2 24.14 24.62 24.68Q3 22.14 22.61 22.61
Competitive in PSNR gain Around ten times faster
Wei FAN 32 / 50I D. Sun and W.-K. Cham, “Postprocessing of low bit-rate block DCT coded images based on a fields of experts
prior,” IEEE TIP, 2007
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
Calibration-Based DCT Histogram Smoothing
J(quality factor q)
IJquality enhancement
IJc
crop by 1 pixel (calibration)
(→ ↓)
Jc
JPEGcom
pression
qualityfactor
q
NqDCT coefficient
subtraction
FJc
DCT coefficientsummation
Calibration:Translation invarianceCropping breaks the 8× 8 block structureNon-parametric DCT quantization noise estimation
Wei FAN 33 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
Histogram Recovery Comparison — FJbq vs. FJ
c
Difference of KL divergence values:
r c 1 2 3 4 5 6 7 81 −0.0081 −0.0001 0.0096 0.0121 0.0075 0.0126 0.0171 −0.02652 0.0037 0.0081 0.0127 0.0025 −0.0013 0.0201 0.0056 −0.03933 0.0171 0.0120 0.0098 −0.0036 −0.0062 −0.0050 −0.0099 −0.06634 0.0133 0.0046 −0.0013 −0.0110 −0.0194 −0.0086 −0.0097 −0.07085 0.0106 0.0022 −0.0037 −0.0159 −0.0291 −0.0237 −0.0591 −0.11736 0.0090 0.0114 −0.0104 −0.0179 −0.0323 −0.0253 −0.0692 −0.12297 0.0353 0.0223 −0.0124 −0.0079 −0.0660 −0.0690 −0.0936 −0.10678 −0.0050 −0.0332 −0.0730 −0.0720 −0.1385 −0.1372 −0.1235 −0.0357
FJc outperforms FJ
bq in low-frequency DCT subbandsThe calibration-based DCT histogram smoothing is faster
Wei FAN 34 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
Proposed JPEG Anti-ForensicsRemove introduced extraunnatural noise
JPEG anti-forensic purposes
Cost Function
arg minu
{
regularization parameter︷︸︸︷λ
2 ‖u− y‖2 +regularization parameter︷︸︸︷
α × ι(u)︸︷︷︸TV of u
+regularization parameter︷︸︸︷
β28∑
k=1
7∑c=0| νk(calibrated image by c pixels︷︸︸︷
uc )︸ ︷︷ ︸variance of subband k
− σ2k︸︷︷︸
estimated variance from FJc
|
+∑
i
regularization parameter︷︸︸︷γ
2 ‖Piu− zi‖2 − log p(auxiliary variable for “Half Quadratic Splitting”︷︸︸︷
zi )}
Wei FAN 35 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
Proposed JPEG Anti-ForensicsRemove introduced extraunnatural noise
JPEG anti-forensic purposes
Cost Function
arg minu
{
regularization parameter︷︸︸︷λ
2 ‖u− y‖2 +regularization parameter︷︸︸︷
α × ι(u)︸︷︷︸TV of u
+regularization parameter︷︸︸︷
β28∑
k=1
7∑c=0| νk(calibrated image by c pixels︷︸︸︷
uc )︸ ︷︷ ︸variance of subband k
− σ2k︸︷︷︸
estimated variance from FJc
|
+∑
i
regularization parameter︷︸︸︷γ
2 ‖Piu− zi‖2 − log p(auxiliary variable for “Half Quadratic Splitting”︷︸︸︷
zi )}
Image fidelity to JPEG image
Wei FAN 35 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
Proposed JPEG Anti-ForensicsRemove introduced extraunnatural noise
JPEG anti-forensic purposes
Cost Function
arg minu
{
regularization parameter︷︸︸︷λ
2 ‖u− y‖2 +regularization parameter︷︸︸︷
α × ι(u)︸︷︷︸TV of u
+regularization parameter︷︸︸︷
β28∑
k=1
7∑c=0| νk(calibrated image by c pixels︷︸︸︷
uc )︸ ︷︷ ︸variance of subband k
− σ2k︸︷︷︸
estimated variance from FJc
|
+∑
i
regularization parameter︷︸︸︷γ
2 ‖Piu− zi‖2 − log p(auxiliary variable for “Half Quadratic Splitting”︷︸︸︷
zi )}
Image fidelity to JPEG imageAnti-forensic terms
Wei FAN 35 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
Proposed JPEG Anti-ForensicsRemove introduced extraunnatural noise
JPEG anti-forensic purposes
Cost Function
arg minu
{
regularization parameter︷︸︸︷λ
2 ‖u− y‖2 +regularization parameter︷︸︸︷
α × ι(u)︸︷︷︸TV of u
+regularization parameter︷︸︸︷
β28∑
k=1
7∑c=0| νk(calibrated image by c pixels︷︸︸︷
uc )︸ ︷︷ ︸variance of subband k
− σ2k︸︷︷︸
estimated variance from FJc
|
+∑
i
regularization parameter︷︸︸︷γ
2 ‖Piu− zi‖2 − log p(auxiliary variable for “Half Quadratic Splitting”︷︸︸︷
zi )}
Image fidelity to JPEG imageAnti-forensic terms
Image prior +“Half Quadratic Splitting”
Wei FAN 35 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
Experimental Results
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KF
KLuo
KQLuo
KV
KL
K1U
K2U
Random guess
FJ0 , TV deblk.
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KF
KLuo
KQLuo
KV
KL
K1U
K2U
Random guess
FJ , TV deblk. + DCT hist. smth.
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KF
KLuo
KQLuo
KV
KL
K1U
K2U
Random guess
FJ1 , GMM + calibration
KF KLuo KQLuo KV KL K1
U K2U PSNR [dB] SSIM
J 0.9991 1.0000 0.9996 0.9976 0.9811 0.9860 0.8840 37.0999 0.9919FJ
Sq Sb0.3783 0.0806 0.6288 0.8337 0.5338 0.6309 0.4854 30.4591 0.9509
IJ 0.9997 0.9982 0.9528 0.7851 0.9698 0.9878 0.8779 37.8930 0.9927FJ
c 0.9994 0.8147 0.5949 0.7383 0.9868 0.9868 0.8944 35.3209 0.9876FJ
1 0.5522 0.7291 0.3594 0.7394 0.5272 0.7750 0.5787 35.2568 0.9832FJ
0 0.6756 0.6046 0.5194 0.6210 0.4490 0.6772 0.5880 35.4814 0.9843FJ 0.5398 0.6425 0.4598 0.6159 0.4344 0.5894 0.5317 35.9855 0.9866
Wei FAN 36 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
Example Results
I, — J , PSNR = 33.3476 dB FJSq Sb
, PSNR = 29.9070 dB FJ1 , PSNR = 32.4401 dB
DCT coefficient value
DCT
coeffi
cientfrequency
−200 −100 0 100 2000
0.04
0.08
0.12
0.16
DCT coefficient value
DCT
coeffi
cientfrequen
cy
−100 −50 0 50 1000
0.1
0.2
0.3
DCT coefficient value
DCT
coeffi
cientfrequency
−30 −20 −10 0 10 20 300
0.1
0.2
0.3
DCT coefficient value
DCT
coeffi
cientfrequency
−10 −5 0 5 100
0.2
0.4
0.6
(2, 2) (1, 6) (7, 4) (8, 8)
Wei FAN 37 / 50I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
3. Using A Sophisticated Image Model
Example Results
I, — J , PSNR = 33.3476 dB FJSq Sb
, PSNR = 29.9070 dB FJ1 , PSNR = 32.4401 dB
DCT coefficient value
DCT
coeffi
cientfrequency
−200 −100 0 100 2000
0.04
0.08
0.12
0.16
DCT coefficient value
DCT
coeffi
cientfrequen
cy
−100 −50 0 50 1000
0.1
0.2
0.3
DCT coefficient value
DCT
coeffi
cientfrequency
−30 −20 −10 0 10 20 300
0.1
0.2
0.3
DCT coefficient value
DCT
coeffi
cientfrequency
−10 −5 0 5 100
0.2
0.4
0.6
(2, 2) (1, 6) (7, 4) (8, 8)
Wei FAN 37 / 50I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.
1050-1065
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
1 Introduction
2 JPEG Anti-Forensics
3 Median Filtering Anti-ForensicsVariational image deconvolution frameworkQuality enhancement & anti-forensics
4 Conclusions & Perspectives
Wei FAN 38 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Median Filtering “Streaking” Artifacts
Smoother local image neighborhood
[84 96 97 8184 85 87 7881 81 88 8584 80 91 84
]
[84 87 87 8982 85 85 8881 84 85 8880 84 84 88
]
Smoother local image neighborhood
[84 96 97 8184 85 87 7881 81 88 8584 80 91 84
]
[84 87 87 8982 85 85 8881 84 85 8880 84 84 88
]
Smoother local image neighborhood
[84 96 97 8184 85 87 7881 81 88 8584 80 91 84
]
[84 87 87 8982 85 85 8881 84 85 8880 84 84 88
]Pixel value difference
Frequency
−200 −100 0 100 2000
0.1
0.2
0.3
Pixel value difference
Frequency
−200 −100 0 100 2000
0.1
0.2
0.3
Smaller pixel value differenceWei FAN 39 / 50I A. C. Bovik. “Streaking in median filtered images”. IEEE TASSP 35, 4 (1987), pp. 493-503
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Image Variational Deconvolution Framework
Median Filtering Process
Convolution Kernel Matrix
Ψk =
ψk1 ψk
4 ψk7
ψk2 ψk
5 ψk8
ψk3 ψk
6 ψk9
, with ψkk = 1 and ψk
i = 0 for ∀i 6= k,i, k ∈ {1, 2, · · · , 9},
Simplification
ΨDBM =
[0.0930 0.1076 0.09270.1109 0.1921 0.11090.0926 0.1074 0.0929
]ΨAVE =
[0.1111 0.1111 0.11110.1111 0.1111 0.11110.1111 0.1111 0.1111
]
ΨGAU =
[0.0113 0.0838 0.01130.0838 0.6193 0.08380.0113 0.0838 0.0113
]
Wei FAN 40 / 50
I H.-D. Yuan. “Blind forensics of median filtering in digital images”. IEEE TIFS 6, 4 (2011), pp. 1335-1345I D. Krishnan, et al. “Blind deconvolution using a normalized sparsity measure”. In: Proc. CVPR. 2011, pp.
233-240
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Image Variational Deconvolution Framework
Pixel Value Difference
Pixel value difference
Frequency
−300 −100 100 3000
0.1
0.2
0.3
Observed dataFitted p.m.f. curve
IPixel value difference
Frequency
−300 −100 100 3000
0.1
0.2
0.3
Observed data
M
Generalized Gaussian Distribution
f (d) = β
2αΓ(1/β)e−(|d|/α)β
Wei FAN 41 / 50I D. Krishnan and R. Fergus. “Fast image deconvolution using hyper-Laplacian priors”. In: Proc. NIPS. 2009, pp.
1033-1041
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Image Variational Deconvolution Framework
Image Variational Deconvolution Framework
Cost Function
arg minu
λ2
median filtering approximation︷ ︸︸ ︷‖Ku− y‖22 +ω ‖u− y‖22︸ ︷︷ ︸
close to MF image
+
histogram regularization︷ ︸︸ ︷∑Jj=1
∥∥∥Fjuαj
∥∥∥βj
βj
Spatially homogenous kernel for approximation
Image fidelity to the median filtered image
Image prior
Wei FAN 42 / 50I D. Krishnan, et al. “Blind deconvolution using a normalized sparsity measure”. In: Proc. CVPR. 2011, pp.
233-240
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Image Variational Deconvolution Framework
Image Variational Deconvolution Framework
Cost Function
arg minu
λ2
median filtering approximation︷ ︸︸ ︷‖Ku− y‖22 +ω ‖u− y‖22︸ ︷︷ ︸
close to MF image
+
histogram regularization︷ ︸︸ ︷∑Jj=1
∥∥∥Fjuαj
∥∥∥βj
βj
Spatially homogenous kernel for approximation
Image fidelity to the median filtered image
Image prior
Wei FAN 42 / 50I D. Krishnan, et al. “Blind deconvolution using a normalized sparsity measure”. In: Proc. CVPR. 2011, pp.
233-240
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Image Variational Deconvolution Framework
Image Variational Deconvolution Framework
Cost Function
arg minu
λ2
median filtering approximation︷ ︸︸ ︷‖Ku− y‖22 +ω ‖u− y‖22︸ ︷︷ ︸
close to MF image
+
histogram regularization︷ ︸︸ ︷∑Jj=1
∥∥∥Fjuαj
∥∥∥βj
βj
Spatially homogenous kernel for approximation
Image fidelity to the median filtered image
Image prior
Wei FAN 42 / 50I D. Krishnan, et al. “Blind deconvolution using a normalized sparsity measure”. In: Proc. CVPR. 2011, pp.
233-240
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Image Variational Deconvolution Framework
Image Variational Deconvolution Framework
Cost Function
arg minu
λ2
median filtering approximation︷ ︸︸ ︷‖Ku− y‖22 +ω ‖u− y‖22︸ ︷︷ ︸
close to MF image
+
histogram regularization︷ ︸︸ ︷∑Jj=1
∥∥∥Fjuαj
∥∥∥βj
βj
Spatially homogenous kernel for approximation
Image fidelity to the median filtered image
Image prior
Wei FAN 42 / 50I D. Krishnan, et al. “Blind deconvolution using a normalized sparsity measure”. In: Proc. CVPR. 2011, pp.
233-240
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Quality Enhancement & Anti-Forensics
Median Filtered Image Quality Enhancement
Original, — Noised (3%)PSNR = 20.8403 dB
Median filteredPSNR = 25.1402 dB
Quality enhancedPSNR = 26.5970 dB
Noise density 1% 3% 5% 7%PSNR [dB] SSIM PSNR [dB] SSIM PSNR [dB] SSIM PSNR [dB] SSIM
Noised 25.1365 0.8308 20.3642 0.6388 18.1466 0.5327 16.6851 0.4632Median filtered 37.1336 0.9827 36.7957 0.9818 36.4031 0.9807 35.8924 0.9793
Quality enhanced 38.0723 0.9892 37.5155 0.9876 36.7914 0.9850 35.7542 0.9803
Parameter setting: ω = 0.4, λ = 1000, γ = 1200
Wei FAN 43 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Quality Enhancement & Anti-Forensics
Median Filtering Anti-Forensics
Pixel Value PerturbationReduce the occurrences of ‘0’ in the first-order pixel differenceMinor image quality sacrifice
Triple Pair
Apply pixel value perturbation before deconvolutionParameter setting: ω = 0.1, λ = 1500, γ = 500
Wei FAN 44 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Quality Enhancement & Anti-Forensics
Experimental Results
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KK
KK
KC
KY
Random guess
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KK
KK
KC
KY
Random guess
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KK
KK
KC
KY
Random guess
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KK
KK
KC
KY
Random guess
M, median filtered FMW , state-of-the-art FM
D , state-of-the-art FM , proposed
Image quality Anti-forensic performance KL divergencePSNR [dB] SSIM KK KK KC KY f1 f2 f3 f4
M 37.2847 0.9831 0.9722 0.9824 0.9938 0.9984 0.1632 0.1611 0.0775 0.0753Mp 38.0580 0.9897 0.7974 0.8587 0.8429 0.8080 0.0925 0.0880 0.0475 0.0437FM
W 33.6033 0.9552 0.4592 0.6586 0.6668 0.3336 0.1148 0.1338 0.0619 0.0689FM
D 33.4272 0.9714 0.5347 0.4635 0.7479 0.6518 0.0547 0.0563 0.0383 0.0389FM 37.5184 0.9901 0.5595 0.5061 0.6490 0.5886 0.0484 0.0449 0.0272 0.0238
Better undetectability Higher image quality (even w.r.t. M)Lower KL divergence values
Wei FAN 45 / 50I Z.-H. Wu, et al. “Anti-forensics of median filtering”. In: ICASSP. 2013, pp. 3043-3047I D. T. Dang-Nguyen, et al. “Counter-forensics of median filtering”. In: MMSP. 2013, pp. 260-265
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Quality Enhancement & Anti-Forensics
Experimental Results
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KK
KK
KC
KY
Random guess
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KK
KK
KC
KY
Random guess
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KK
KK
KC
KY
Random guess
False positive rate
Truepositiverate
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
KK
KK
KC
KY
Random guess
M, median filtered FMW , state-of-the-art FM
D , state-of-the-art FM , proposed
Image quality Anti-forensic performance KL divergencePSNR [dB] SSIM KK KK KC KY f1 f2 f3 f4
M 37.2847 0.9831 0.9722 0.9824 0.9938 0.9984 0.1632 0.1611 0.0775 0.0753Mp 38.0580 0.9897 0.7974 0.8587 0.8429 0.8080 0.0925 0.0880 0.0475 0.0437FM
W 33.6033 0.9552 0.4592 0.6586 0.6668 0.3336 0.1148 0.1338 0.0619 0.0689FM
D 33.4272 0.9714 0.5347 0.4635 0.7479 0.6518 0.0547 0.0563 0.0383 0.0389FM 37.5184 0.9901 0.5595 0.5061 0.6490 0.5886 0.0484 0.0449 0.0272 0.0238
Better undetectability Higher image quality (even w.r.t. M)Lower KL divergence values
Wei FAN 45 / 50I Z.-H. Wu, et al. “Anti-forensics of median filtering”. In: ICASSP. 2013, pp. 3043-3047I D. T. Dang-Nguyen, et al. “Counter-forensics of median filtering”. In: MMSP. 2013, pp. 260-265
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Quality Enhancement & Anti-Forensics
Against SVM-Based Detectors
Replacement rate
AUC
0 0.1 0.2 0.3 0.4 0.5 0.6 0.8 10.5
0.6
0.7
0.8
0.9
1
M
FM
W
FM
D
FM
Replacement rate
AUC
0 0.1 0.2 0.3 0.4 0.5 0.6 0.8 10.5
0.6
0.7
0.8
0.9
1
M
FM
W
FM
D
FMKS686SPAM KS44
MFF
Replacement rate
AUC
0 0.1 0.2 0.3 0.4 0.5 0.6 0.8 10.5
0.6
0.7
0.8
0.9
1
M
FM
W
FM
D
FM
Replacement rate
AUC
0 0.1 0.2 0.3 0.4 0.5 0.6 0.8 10.5
0.6
0.7
0.8
0.9
1
M
FM
W
FM
D
FM
Replacement rate
AUC
0 0.1 0.2 0.3 0.4 0.5 0.6 0.8 10.5
0.6
0.7
0.8
0.9
1
M
FM
W
FM
D
FMKS56GLF KS10
AR KS220LTP
FM FM
FM FM FM
Wei FAN 46 / 50
I T. Pevny, et al. “Steganalysis by subtractive pixel adjacency matrix”. IEEE TIFS 5, 2 (2010), pp. 215-224I H.-D. Yuan. “Blind forensics of median filtering in digital images”. IEEE TIFS 6, 4 (2011), pp. 1335-1345I C. Chen, et al. “Blind detection of median filtering in digital images: a difference domain based approach”. IEEE
TIFS 22, 12 (2013), pp. 4699-4710I X. Kang, et al. “Robust median filtering forensics using an autoregressive model”. IEEE TIFS 8, 9 (2013), pp.
1456-1468I Y. Zhang, et al. “Revealing the traces of median filtering using high-order local ternary patterns”. IEEE Signal
Processing Letters 21, 3 (2014), pp. 275-280
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Quality Enhancement & Anti-Forensics
Application: JPEG Deblocking
Image quality Anti-forensic performancePSNR [dB] SSIM KF K1
U K2U KK KK KC KY
J 42.9742 0.9975 0.9930 0.9871 0.8937 0.3895 0.2699 0.3990 0.4985Jm 37.0888 0.9830 0.6128 0.7151 0.5288 0.9659 0.9786 0.9924 0.9986J w 33.5111 0.9549 0.6321 0.5709 0.5363 0.4109 0.6061 0.6369 0.3225J d 33.3101 0.9711 0.4945 0.6255 0.5274 0.5035 0.4256 0.7230 0.6489J f 37.2413 0.9894 0.6003 0.5168 0.4754 0.5137 0.4489 0.6244 0.5976
High image qualityGood forensic undetectability against both JPEG blocking / medianfiltering forensic detectors
Wei FAN 47 / 50
I M. C. Stamm and K. J. R. Liu. “Anti-forensics of digital image compression”. IEEE TIFS 6, 3 (2011), pp.1050-1065
I Z.-H. Wu, et al. “Anti-forensics of median filtering”. In: ICASSP. 2013, pp. 3043-3047I D. T. Dang-Nguyen, et al. “Counter-forensics of median filtering”. In: MMSP. 2013, pp. 260-265
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Conclusions
1 A new research line of designing imageanti-forensics via image restoration
2 JPEG anti-forensics using TV-baseddeblocking
3 Perceptual DCT histogram smoothing
4 Leveraging on a sophisticated imageprior model and calibration:
JPEG image quality enhancement
Non-parametric DCT quantizationnoise estimation
5 Median filtered image quality en-hancement and anti-forensics via vari-ational deconvolution
I Wei Fan, Kai Wang, Francois Cayre, and Zhang Xiong,“A variational approach to JPEG anti-forensics”, InProc. of the IEEE International Conference on Acous-tics, Speech, and Signal Processing (ICASSP), Vancou-ver, Canada, pp. 3058-3062, 2013.
I Wei Fan, Kai Wang, Francois Cayre, and Zhang Xiong,“JPEG anti-forensics with improved tradeoff betweenforensic undetectability and image quality”, IEEE Trans-actions on Information Forensics and Security, vol. 9,no. 8, pp. 1211-1226, 2014.
I Wei Fan, Kai Wang, Francois Cayre, and Zhang Xiong,“JPEG anti-forensics using non-parametric DCT quanti-zation noise estimation and natural image statistics”, InProc. of the ACM International Workshop on Informa-tion Hiding and Multimedia Security (ACM IHMMSec),Montpellier, France, pp. 117-122, 2013. (Best PaperAward)
I Wei Fan, Kai Wang, Francois Cayre, and Zhang Xiong,“Median filtered image quality enhancement and anti-forensics via variational deconvolution”, IEEE Transac-tions on Information Forensics and Security, vol. 10, no.5, pp. 1076-1091, 2015.
Wei FAN 48 / 50
I Wei Fan, Kai Wang, Francois Cayre, and Zhang Xiong, “3-D lighting-based image forgery detection using shape-from-shading”, In Proc. of the European Signal Processing Conference (EUSIPCO), Bucharest, Romania, IEEE,pp. 1777-1781, 2012.
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Conclusions
1 A new research line of designing imageanti-forensics via image restoration
2 JPEG anti-forensics using TV-baseddeblocking
3 Perceptual DCT histogram smoothing
4 Leveraging on a sophisticated imageprior model and calibration:
JPEG image quality enhancement
Non-parametric DCT quantizationnoise estimation
5 Median filtered image quality en-hancement and anti-forensics via vari-ational deconvolution
I Wei Fan, Kai Wang, Francois Cayre, and Zhang Xiong,“A variational approach to JPEG anti-forensics”, InProc. of the IEEE International Conference on Acous-tics, Speech, and Signal Processing (ICASSP), Vancou-ver, Canada, pp. 3058-3062, 2013.
I Wei Fan, Kai Wang, Francois Cayre, and Zhang Xiong,“JPEG anti-forensics with improved tradeoff betweenforensic undetectability and image quality”, IEEE Trans-actions on Information Forensics and Security, vol. 9,no. 8, pp. 1211-1226, 2014.
I Wei Fan, Kai Wang, Francois Cayre, and Zhang Xiong,“JPEG anti-forensics using non-parametric DCT quanti-zation noise estimation and natural image statistics”, InProc. of the ACM International Workshop on Informa-tion Hiding and Multimedia Security (ACM IHMMSec),Montpellier, France, pp. 117-122, 2013. (Best PaperAward)
I Wei Fan, Kai Wang, Francois Cayre, and Zhang Xiong,“Median filtered image quality enhancement and anti-forensics via variational deconvolution”, IEEE Transac-tions on Information Forensics and Security, vol. 10, no.5, pp. 1076-1091, 2015.
Lessons Learned
1 Better image anti-forensicmethods are designed
2 Current forensic methodsare not that reliable
3 Natural image statistics isimportant
Wei FAN 48 / 50
I Wei Fan, Kai Wang, Francois Cayre, and Zhang Xiong, “3-D lighting-based image forgery detection using shape-from-shading”, In Proc. of the European Signal Processing Conference (EUSIPCO), Bucharest, Romania, IEEE,pp. 1777-1781, 2012.
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Perspectives
Short Term
Anti-forensic JPEG image with better quality than the JPEG image
Other image anti-forensic methods leveraging on image restoration
Anti-forensics on other digital media, e.g., video
Long Term
Universal image anti-forensics
Open Questions
A single step attack for JPEG anti-forensics?Estimation of the spatially heterogeneous convolution kernel for me-dian filtering?Anti-forensic image, as a whole, against machine learning basedforensic detectors?
Wei FAN 49 / 50
Introduction JPEG Anti-Forensics Median Filtering Anti-Forensics Conclusions & Perspectives
Thank you for your attention!
Q & A
Wei FAN 50 / 50