forgery manipulation detection: challenges and...

68
WVU, Anchorage - 2008 .:. Rocha & Goldenstein, Forgery manipulation detection Forgery manipulation detection: challenges and trends Institute of Computing University of Campinas (Unicamp) CEP 13084-851, Campinas, SP - Brazil Siome Goldenstein [email protected] Anderson Rocha [email protected] 1

Upload: others

Post on 10-Oct-2019

9 views

Category:

Documents


2 download

TRANSCRIPT

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Forgery manipulation detection:challenges and trends

Institute of ComputingUniversity of Campinas (Unicamp)

CEP 13084-851, Campinas, SP - Brazil

Siome [email protected]

Anderson [email protected]

1

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection 2

Can we trust images?

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Forgery scenario

Digital Forensics Analysis

5

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

‣ Introduction

‣ Terminology

‣ Historical aspects

‣ Techniques

‣ Opportunities

Summary

6

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Introduction

7

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Introduction

‣ What is Digital Image Forensics?

‣ Motivation

• Crime judgement

• Proof destruction

• Creation/forgery of events

8

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Historical aspects

9

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Two ways of life by Oscar Rejland, 1857.

Historical aspects

10

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Historical aspects

Stalin with (original) and without (doctored) Nikolai Yezhov.

11

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Historical aspects

12

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Historical aspects

Israeli attack on Lebanon. Adnan Hajj photographer

darkened and dramatized the event.

12

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Historical aspects

13

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Historical aspects

US soldier “guides” an Iraqi with his child.

Photograph and forgery by Brian Walski.

13

One of the most impressive news photos of 2006

Liu Weiqiang of the Daqing Evening News.

Historical aspects

More doctoring examples.

Historical aspects

More doctoring examples.

Historical aspects

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Recently...

17

Recent discoveries

Rome protest, 2004.

D. Sacchi, F. Agnoli, E. Loftus. Applied Cognitive Psychology, vol. 21, n. 8, 249-273, 2007.

Recent discoveriesCredits to Stwart Franklin, 1989

D. Sacchi, F. Agnoli, E. Loftus. Applied Cognitive Psychology, vol. 21, n. 8, 249-273, 2007.

Beijing, 1989.

Science frauds.

Recent discoveries

(a) Erasing (b) Removing(c) Replicating

Top right: healing. Bottom: texture maps

H. Farid. Exposing Digital Forgeries in Scientific Images. ACM Multimedia and Security Workshop, 2006.

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Techniques

‣ Composition

‣ Retouching

‣ Sharpening

‣ Computer generation

21

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

‣ Is this image an “original” image or was it created by means of composition (copy/paste)?

‣ Does this image represent a trully scene/event or was it digitally tampered to deceive the viewer?

Important questions

22

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

‣ What is the processing history of this image?

‣ What parts of the image has undergone any kind of processing and up to what extent?

‣ Was the image acquired by a source manufactured by vendor X or Y?

Important questions

23

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

‣ Passive blind image analysis

‣ No watermarking needed at all

24

Community efforts

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

‣ Source identification

‣ Computer generated images identification

‣ Forgery detection

Three branches

25

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Forgery detection branches

‣ Approaches based on variations of some image features

‣ Approaches based on image features inconsistencies

‣ Approaches based on image acquisition process inconsistencies

26

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

General camera pipelineLight Lens System

Exposure, focusing and image stabilization

Filters

Infre-red, anti-aliasing... for max. visible quality

Imaging sensors

CCD, CMOS...

Color Filter Arrays (CFA)...

Mosaicing

• Demosaicing• White point

correction, • Image sharpening• Aperture correction• Gamma correction• Compression...

DIP

Resulting photograph

27

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

‣ Ng et al. studied the effects of image splicing on magnitude and phase characteristics of the normalized bispectrum (bicoherence)

Normalized bispectrum is the Fourier transform of the third moment of a signal

‣ ~62% and high cost implementation

Variations in image features

28T. T. Ng, S. Chang, and Q. Sun. Blind detection of photomontage using higher order statistics. ISCAS, 2004

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Variations in image features

‣ Avcibas et al. proposed to use IQMs

‣ The approach calculates deviations between the image under analysis and its estimated original version (obtained through denoising)

I. Avcibas, S. Bayram, N. Memon, B. Sankur, M. Ramkumar. A classifier design for detecting image manipulations. IEEE ICIP, 2004. 29

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Variations in image features

‣ Analysis for controlled manipulations (brighteness, scaling, blurring, sharpening, rotation)

‣ ~74% accuracy

‣ Bayram et al. proposed the inclusion of BSMs and wavelets coeficient analysis

‣ ~90% accuracy

S. Bayram, I. Avcibas, N. Memon, B. Sankur. Image manipulation detection. JEI, vol. 15, no. 4, 2006. 30

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

‣ Farid et al. and also Fridrich et al. in independent projects have analyzed JPEG recompression artifacts

‣ They found that recompression of an (already compressed) image at a different quality factor distorts the smoothness of DCT coefficient histograms

Image features inconsistencies

31

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

‣ Essentially, the commonest image tampering approach involves splicing of images with different compression levels (therefore able to be detected)

‣ However, double compression is not a proof of tampering

Image features inconsistencies

32

Histogram of a normally distributed signal

Histogram of a single quantized signal (Step 2)

Histogram of a double quantized signal (Steps 3 followed by 2)

Histogram of a normally distributed signal

Histogram of a single quantized signal (Step 3)

Histogram of a double quantized signal (Steps 2 followed by 3)

Image features inconsistencies

Image features inconsistencies

‣What are the problems with the JPEG double quantization detection?

Image features inconsistencies

‣What are the problems with the JPEG double quantization detection?

• Cropping

Image features inconsistencies

‣What are the problems with the JPEG double quantization detection?

• Cropping

• High-quality compression followed by significantly lower quality compression

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

‣ Popescu and Farid (2005), proposed a method for detecting traces of resampling

‣ The principle is that upsampling (interpolation) introduces periodic inter-coefficients correlations

‣ Resampling at arbitrary rates require combinations of up-sampling and down-sampling operations

Image features inconsistencies

36

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Image features inconsistencies

Mosaicing

Demosaicing

37

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

‣ They used an EM algorithm to estimate the distribution parameters

‣ ~100% accuracy for RAW images

‣ Accuracy drops down under JPEG compression

Image features inconsistencies

38

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Bilinear

Bi-cubic

Smooth Hue

No CFAinterpolation

Image Prob. map p |F(p)|

Estimated interpolation coefficients from 100 images CFA interpolated with eight different algorithms.

39

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

OriginalTampered

Tampered Original

40

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

41

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

‣What are the problems with color filter array analysis?

41

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

‣What are the problems with color filter array analysis?

• Compression

41

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

‣What are the problems with color filter array analysis?

• Compression

• You can discover the demosaicing algorithm and apply it after the modifications

41

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Image features inconsistencies

‣ Repetition of image parts is a common form of forgery

‣ This kind of tampering can be easily detected with exaustive search and analysis of correlation parts

‣ Exaustive search-based methods are not computationally practical

42

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Image features inconsistencies

‣ Fridrich et al. (2003) proposed a faster (nlogn) and more accurate method

‣ The methods obtains DCT coefficients from overlapping sliding windows

‣ The resulting coefficients are disposed row-wise in a matrix and lexicographically sorted

J. Fridrich, D. Soukal, and J. Lukas. Detection of copy-move forgery in digital images. DFRWS, 2003. 43

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

‣ Correlated blocks under a specified threshold are tagged as duplicated

‣ A similar approach was proposed by Popescu and Farid (2005) that usesPCA coefficients instead of DCT

Image features inconsistencies

A. Popescu and H. Farid. Exposing Digital Forgeries by Detecting Duplicated Image Regions.TR 2004-515, 2004. 44

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Image features inconsistencies

A. Popescu and H. Farid. Exposing Digital Forgeries by Detecting Duplicated Image Regions.TR 2004-515, 2004.

Original Doctored Duplication maps

45

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

A. Swaminathan, M. Wu, and K. Liu. Image tampering identification using blind deconvolution. ICIP, 2006.

‣ Swaminathan et al. used inconsistencies in color filter array interpolation

‣ After estimating the CFA pattern and the interpolation filter, the demosaiced image is reconstructed and compared to the image itself

46

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

J. Lukas, J. Fridrich, and M. Goljan. Detecting digital image forgeries using sensor pattern noise. Proc. of SPIE, 2006.

‣ Lukas et al. (2006) proposed to use inconsistencies in the sensor pattern noise extracted from an image

‣ The noise patterns obtained from various regions are correlated with the corresponding regions in the camera’s reference pattern and a decision is made

47

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

48

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

‣What is the main problem with sensor pattern noise based approaches?

48

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

‣What is the main problem with sensor pattern noise based approaches?

• Flatfielding. It analyzes the two main image noise components (FPN, PRNU)

48

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

‣What is the main problem with sensor pattern noise based approaches?

• Flatfielding. It analyzes the two main image noise components (FPN, PRNU)

• FPN estimated with dark frame

48

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

‣What is the main problem with sensor pattern noise based approaches?

• Flatfielding. It analyzes the two main image noise components (FPN, PRNU)

• FPN estimated with dark frame

• PRNU needs L images with homogeneously illuminated scene

48

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

M. Jonhson and H. Farid. Exposing Digital Forgeries by Detecting Inconsistencies in Lighting. ACM Multimedia and Security Workshop, 2005

‣ Johnson and Farid (2005) analyzed the inconsistencies in light direction

‣ To estimate the light directions, the authors assume

• Surface is Lambertian (reflects light isotropically)

• Patches have a constant reflectance value (as opposite to the entire surface)

• It is illuminated by a point light source infinitelly far away

49

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

‣ Common solutions to estimate point light sources require knowledge of the 3-D surface normals and, at least, four distinct points of the surface with the same reflectance

‣ That would require more than one image or a known object in the scene (e.g., a sphere)

‣ For forensic applications, such solutions are not practical

50

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

P. Nillius and J. Eklundh. Automatic estimation of the projected light source direction. CVPR, 2001

‣ Instead, the authors used an approach proposed by Nillius and Eklundh (2001) to calculate a single directional light source from only one image

‣ That makes sense for outdoor images

‣ Authors also present an extension for more than one point light source (indoor images)

51

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Inconsistencies in the acquisition

Diffuse non-directional light

Directional light

~123º

~86º

~98º

~98º

~93º

52

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Challenges

‣ Performance evaluation and benchmarking

‣ Current approaches mainly present proofs of concept

‣ Proper data sets need to be designed and shared

‣ Robustness issues -- attacks barely studied in the literature

53

WVU, Anchorage - 2008 .:.Rocha & Goldenstein, Forgery manipulation detection

Questions?

The thinker by Rodin

55