Ranking of manipulated images in a large set using Error Level - PowerPoint PPT Presentation

12-02-12 Ranking of manipulated images in a large set using Error Level Analysis Daan Wagenaar & Jeffrey Bosma University of Amsterdam In cooperation with the Netherlands Forensic Institute

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Agenda ¤ Image Manipulation ¤ Research Question ¤ Error Level Analysis ¤ Methodology ¤ Experiments ¤ Results ¤ Conclusion ¤ Further Research ¤ Questions 2

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Agenda ¤ Image Manipulation ¤ Research Question ¤ Error Level Analysis ¤ Methodology ¤ Experiments ¤ Results ¤ Conclusion ¤ Further Research ¤ Questions 3

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Image Manipulation ¤ Examples ¤ Red Eye removal ¤ Brightness enhancements ¤ Sharpening ¤ … ¤ Most interesting manipulations ¤ Internal copy & move ¤ External copy & move 4

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Object removal Stalin without Yezhov Stalin with Yezhov (manipulated) (original) 5

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Object appearance modification Katie Couric Slimmed Body (original) (manipulated) 6

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Object addition Holding an iPhone Holding a BlackBerry (original) (manipulated) 7

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Agenda ¤ Image Manipulation ¤ Research Question ¤ Error Level Analysis ¤ Methodology ¤ Experiments ¤ Results ¤ Conclusion ¤ Further Research ¤ Questions 8

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Research Question ¤ Problem: ¤ A set of images as part of evidence ¤ An expert manually inspects each image for manipulations ¤ Time consuming process in a large set of images v Can the Error Level Analysis technique be used to rank a set of images according to potentially present image manipulation? 9

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Agenda ¤ Image Manipulation ¤ Research Question ¤ Error Level Analysis ¤ Methodology ¤ Experiments ¤ Results ¤ Conclusion ¤ Further Research ¤ Questions 10

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Error Level Analysis (ELA) ¤ A technique for detecting image manipulations ¤ Uses properties of lossy image format ¤ Compares error caused by compression to a certain quality level ¤ An example: ¤ Initial image at a quality level of 95% ¤ ELA resaves this image at a certain quality level (e.g. 95%) ¤ Compression introduces error ¤ Compare error of initial and resaved image ¤ Manipulated areas will have a different level of error ¤ Differences are visibly expressed by brightness in a third image 11

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 ELA @75% Original image ELA @ 85% ELA @ 95% 12

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Manipulated image ELA @75% ELA @ 85% ELA @ 95% 13

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 ¤ Limitations ¤ False positives can be caused by: ¤ Sharp contrast, well-defined patterns ¤ Recoloring, such as brightening, pallet skew, ... ¤ False negatives can be caused by: ¤ Low resolutions ¤ Scaling ¤ Low quality ¤ Image scanning from other sources ¤ Extremely skilled artists 14

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Agenda ¤ Image Manipulation ¤ Research Question ¤ Error Level Analysis ¤ Methodology ¤ Experiments ¤ Results ¤ Conclusion ¤ Further Research ¤ Questions 15

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Methodology ¤ Method 1: Average RGB values per block ¤ Method 2: Block to block comparison ¤ Method 3: Colored pixels ratio ¤ Method 4: Highest luminance value of the brightest pixel ¤ Method 5: Average luminance value of the 64 brightest pixels ¤ Method 6: Average luminance value of the brightest block 16

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Agenda ¤ Image Manipulation ¤ Research Question ¤ Error Level Analysis ¤ Methodology ¤ Experiments ¤ Results ¤ Conclusion ¤ Further Research ¤ Questions 17

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Experiments ¤ Goal ¤ Proof of concept ¤ Dataset of 300 images ¤ 100 images with Canon PowerShot A630 ¤ 100 images with iPhone 4 ¤ 100 images with Samsung Digimax S500 ¤ 30 manipulated images 18

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Agenda ¤ Image Manipulation ¤ Research Question ¤ Error Level Analysis ¤ Methodology ¤ Experiments ¤ Results ¤ Conclusion ¤ Further Research ¤ Questions 19

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Results Rankings with ELA at 75% 100 Method 3 (Colored pixels ratio) Method 4 (Highest luminance value of the brightest pixel) Method 5 (Average luminance value of the 64 brightest pixels) Method 6 (Average luminance value of the brightest block) 80 Manipulated images found (%) 60 40 20 0 0 50 100 150 200 250 300 Rank 20

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Rankings with ELA at 85% 100 Method 3 (Colored pixels ratio) Method 4 (Highest luminance value of the brightest pixel) Method 5 (Average luminance value of the 64 brightest pixels) Method 6 (Average luminance value of the brightest block) 80 Manipulated images found (%) 60 40 20 0 0 50 100 150 200 250 300 Rank 21

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Rankings with ELA at 95% 100 Method 3 (Colored pixels ratio) Method 4 (Highest luminance value of the brightest pixel) Method 5 (Average luminance value of the 64 brightest pixels) Method 6 (Average luminance value of the brightest block) 80 Manipulated images found (%) 60 40 20 0 0 50 100 150 200 250 300 Rank 22

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Manipulated image ELA @75% ELA @ 85% ELA @ 95% 23

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Manipulated image ELA @75% ELA @ 85% ELA @ 95% 24

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Manipulated image ELA @75% ELA @ 95% ELA @ 85% 25

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Original image ELA @75% ELA @ 85% ELA @ 95% 26

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Manipulated image ELA @75% ELA @ 85% ELA @ 95% 27

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Agenda ¤ Image Manipulation ¤ Research Question ¤ Error Level Analysis ¤ Methodology ¤ Experiments ¤ Results ¤ Conclusion ¤ Further Research ¤ Questions 28

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Conclusion ¤ Most effective method ¤ Limitations of ELA directly affect developed methods ¤ Detectable manipulation techniques v Can the Error Level Analysis technique be used to rank a set of images according to potentially present image manipulation? ¤ Yes, it is possible albeit not very reliable. 29

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Agenda ¤ Image Manipulation ¤ Research Question ¤ Error Level Analysis ¤ Methodology ¤ Experiments ¤ Results ¤ Conclusion ¤ Further Research ¤ Questions 30

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Further Research ¤ Alternative to ELA ¤ Combine different rankings ¤ Different methods 31

Ranking of manipulated images in a large set using Error Level Analysis 12-02-12 Questions? 32