Instruction for experiments

The experiments are conducted based on Sarkar et al.�s approach. The starting images are prepared as concluded below. For example for the experiment for compressed images at QF100:

1. Compress the images in databases to JPEG at QF75, by using code : ( click at filename to download )

- comp1.m – to compress the UCID images in folder �UCID1338� to be 512×384 pixels JPEG QF75 for folder �UCIDQF75�, the file names are in number run from 1 to 1338.jpg

- comp2.m – to compress the UCUS images in folder �UCUS1009� to be 512×384 pixels JPEG QF75 for folder �UCUSQF75�, the file names are in number run from 1 to 1009 .jpg

2. Prepare decompressed images before seam modifications.

- decomp1.m – to decompress the UCID images in �UCIDQF75� for folder �UCIDQ75TIF�, and ready for retargeting the images.

- decomp2.m – to decompress the UCUS images in �UCUSQF75� for folder �UCUSQ75TIF�, and ready for retargeting the images.

3. Retarget the images in step 2 by using seam modifications for each different tamper rates and classify into different dataset namely, 1%, 2%, 5%, 10%, 20%, 30%, 50% and mixed set.

v For example, retargeting the UCID images with 1% seam carving; the needed codes are:

- doEdge.m called by mQF100byUCIDQ75TIFC01.m

- doDownEnergy.m called by mQF100byUCIDQ75TIFC01.m

- mQF100byUCIDQ75TIFC01.m – to retarget the images by 1% seam carving rate. This code can be applied for others seam carving rate by changing the operand of 1-0.99 in line 15.

Line 14: % Seam carving rate 1% = 1-0.99 retargeting

Line 15: cut_num = floor(n*(1-0.99));

For instance,

2% seam carving rate must change the operand 0.99 to be 0.98,

5% seam carving rate must change the operand 0.99 to be 0.95,

10% seam carving rate must change the operand 0.99 to be 0.9,

20% seam carving rate must change the operand 0.99 to be 0.8,

30% seam carving rate must change the operand 0.99 to be 0.7, and

50% seam carving rate must change the operand 0.99 to be 0.5.

The images in target path �QF100forUCIDQ75TIFC01� are named by number run from 1339 to 2676 .jpg and resulted at QF100.

Line 66: % File names are resulted in number run from 1339 to 2676 .jpg

Line 67: outName= ['./QF100byUCIDQ75TIFC01/',num2str(pic+1338),'.jpg'];

v For example, retargeting the UCUS images with 1% seam insertion; the needed codes are:

- doEdge2.m called by mQF100forUCIDQ75TIFI01.m

- doDownEnergy.m called by mQF100forUCIDQ75TIFI01.m

- mQF100byUCIDQ75TIFI01.m – to retarget the images by 1% seam insertion rate. This code can be applied for others seam insertion rate by changing the operand of n_o * 0.01 in line 16.

Line 15: % Seam insertion rate 1% = n_o * 0.01 retargeting

Line 16: cut_num = floor(n_o * 0.01);

For instance,

2% seam carving rate must change the operand 0.01 to be 0.02,

5% seam carving rate must change the operand 0.01 to be 0.05,

10% seam carving rate must change the operand 0.01 to be 0.1,

20% seam carving rate must change the operand 0.01 to be 0.2,

30% seam carving rate must change the operand 0.01 to be 0.3, and

50% seam carving rate must change the operand 0.01 to be 0.5.

The images in target path �QF100forUCUSQ75TIFI01� are named by number run from 1010 to 2018 .jpg and resulted at QF100.

Line 106: % File names are resulted in number run from 1010 to 2018 .jpg

Line 107: outName= ['./QF100byUCUSQ75TIFI01/',num2str(pic+1009),'.jpg'];

4. Compress the retargeted images to be JPEG at QF100 before passing through the detection process.

- By mQF100forUCIDQ75TIFC01.m, can be found at

Line 70: % Compressed rate at Quality 100 --> QF100

Line 71: imwrite(targetImg, outName, 'Quality', 100);

- By mQF100forUCIDQ75TIFI01.m, can be found at

Line 109: % Compressed rate at Quality 100 --> QF100

Line 110: imwrite(targetImg, outName, 'Quality', 100);

5. Calculate the matrix to show the symmetry phenomena in blocking effects.

5.1. Read the JPEG images obtained by compressing at QF75 (the original images dataset), and the images obtained by seam modifications (the tamper images dataset) into the calculation process by using code:

- BACMQF100UCIDC01.m, is prepared to read the data from the original images data and tamper images datasets.

Line 4: % How many images you want to analysis

Line 5: numPic = 1338+1338;

�

Line 15: % The images' paths

Line 16: % For the original images and tamper images dataset

Line 17: if pic < 1339;

Line 18: sourceName= ['./UCIDQF75/',num2str(pic),'.jpg'];

Line 19: else

Line 20: sourceName= ['./QF100byUCIDQ75TIFC01/',num2str(pic),'.jpg'];

Line 21: end

Line 22: img= imread(sourceName);

�

5.2. Analyze the BACM data and output as .txt file.

Line 258: % The analyzed data output

Line 259: [fid,message] = fopen('BACMQ100UCIDC01.txt','w+');

�

Line 263: % How many modified images in the dataset

Line 264: if i>1338

�

6. Train the BACM to create a training model for cross validation test. In our example here, we train the data by

- svm-train.exe command in libsvm-3.14 or libsvm-3.17 package. The command is found in subfolder windows (libsvm-3.14/windows). Train the �BACMQF100UCIDC01.txt� to create �BACMQ100UCIDC01.txt.model�.

- For instance of the command lines for training, and the results were captured and demonstrated below:

Phoenix:windows Phoenix$ svm-train BACMQ100UCIDC01.txt

..*.*

optimization finished, #iter = 3390

nu = 0.376003

obj = -726.115364, rho = 0.307937

nSV = 1696, nBSV = 576

Total nSV = 1696




Phoenix:windows Phoenix$ svm-train BACMQ100UCIDC02.txt

..*.*

optimization finished, #iter = 3410

nu = 0.338551

obj = -653.582775, rho = 0.259299

nSV = 1617, nBSV = 476

Total nSV = 1617

Phoenix:windows Phoenix$ svm-train BACMQ100UCIDC10.txt

.*.*

optimization finished, #iter = 2396

nu = 0.178951

obj = -311.840801, rho = -0.153669

nSV = 1044, nBSV = 187

Total nSV = 1044




Phoenix:windows Phoenix$ svm-train BACMQ100UCIDC20.txt

.*

optimization finished, #iter = 1900

nu = 0.108905

obj = -183.776140, rho = -0.477221

nSV = 812, nBSV = 117

Total nSV = 812




Phoenix:windows Phoenix$ svm-train BACMQ100UCIDC30.txt

*.*

optimization finished, #iter = 1488

nu = 0.063856

obj = -108.088860, rho = -0.698763

nSV = 683, nBSV = 74

Total nSV = 683




Phoenix:windows Phoenix$ svm-train BACMQ100UCIDC50.txt

*.*

optimization finished, #iter = 1090

nu = 0.025790

obj = -45.448209, rho = -0.895734

nSV = 575, nBSV = 34

Total nSV = 575

Phoenix:windows Phoenix$ svm-train BACMQ100UCIDMixed7C.txt

.*.*

optimization finished, #iter = 2763

nu = 0.236235

obj = -423.774780, rho = 0.069050

nSV = 1216, nBSV = 252

Total nSV = 1216

7. Cross validation test by

- svm-predict.exe command in libsvm-3.14 or libsvm-3.17 package. The command is found in subfolder windows (libsvm-3.14/windows).

- For instance of the command lines for cross validation test, and the results were captured and demonstrated below:

% cross validation test that tested by BACMQ100UCIDC01.txt.model for all test dataset

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC01.txt BACMQ100UCIDC01.txt.model testQ100UCIDC01TC01.txt

Accuracy = 92.1151% (2465/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC02.txt BACMQ100UCIDC01.txt.model testQ100UCIDC01TC02.txt

Accuracy = 92.2646% (2469/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC05.txt BACMQ100UCIDC01.txt.model testQ100UCIDC01TC05.txt

Accuracy = 95.0673% (2544/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC10.txt BACMQ100UCIDC01.txt.model testQ100UCIDC01TC10.txt

Accuracy = 96.8236% (2591/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC20.txt BACMQ100UCIDC01.txt.model testQ100UCIDC01TC20.txt

Accuracy = 97.7952% (2617/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC30.txt BACMQ100UCIDC01.txt.model testQ100UCIDC01TC30.txt

Accuracy = 98.1689% (2627/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC50.txt BACMQ100UCIDC01.txt.model testQ100UCIDC01TC50.txt

Accuracy = 98.2063% (2628/2676) (classification)




% cross validation test that tested by BACMQ100UCIDC02.txt.model for all test dataset

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC01.txt BACMQ100UCIDC02.txt.model testQ100UCIDC02TC01.txt

Accuracy = 89.574% (2397/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC02.txt BACMQ100UCIDC02.txt.model testQ100UCIDC02TC02.txt

Accuracy = 92.9372% (2487/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC05.txt BACMQ100UCIDC02.txt.model testQ100UCIDC02TC05.txt

Accuracy = 95.0299% (2543/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC10.txt BACMQ100UCIDC02.txt.model testQ100UCIDC02TC10.txt

Accuracy = 96.9731% (2595/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC20.txt BACMQ100UCIDC02.txt.model testQ100UCIDC02TC20.txt

Accuracy = 97.9447% (2621/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC30.txt BACMQ100UCIDC02.txt.model testQ100UCIDC02TC30.txt

Accuracy = 98.3931% (2633/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC50.txt BACMQ100UCIDC02.txt.model testQ100UCIDC02TC50.txt

Accuracy = 98.4305% (2634/2676) (classification)

% cross validation test that tested by BACMQ100UCIDC05.txt.model for all test dataset




Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC01.txt BACMQ100UCIDC05.txt.model testQ100UCIDC05TC01.txt

Accuracy = 88.1166% (2358/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC02.txt BACMQ100UCIDC05.txt.model testQ100UCIDC05TC02.txt

Accuracy = 90.5082% (2422/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC05.txt BACMQ100UCIDC05.txt.model testQ100UCIDC05TC05.txt

Accuracy = 95.6652% (2560/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC10.txt BACMQ100UCIDC05.txt.model testQ100UCIDC05TC10.txt

Accuracy = 96.9731% (2595/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC20.txt BACMQ100UCIDC05.txt.model testQ100UCIDC05TC20.txt

Accuracy = 98.281% (2630/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC30.txt BACMQ100UCIDC05.txt.model testQ100UCIDC05TC30.txt

Accuracy = 98.58% (2638/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC50.txt BACMQ100UCIDC05.txt.model testQ100UCIDC05TC50.txt

Accuracy = 98.6547% (2640/2676) (classification)

% cross validation test that tested by BACMQ100UCIDC10.txt.model for all test dataset

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC01.txt BACMQ100UCIDC10.txt.model testQ100UCIDC10TC01.txt

Accuracy = 79.5217% (2128/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC02.txt BACMQ100UCIDC10.txt.model testQ100UCIDC10TC02.txt

Accuracy = 84.3423% (2257/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC05.txt BACMQ100UCIDC10.txt.model testQ100UCIDC10TC05.txt

Accuracy = 91.704% (2454/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC10.txt BACMQ100UCIDC10.txt.model testQ100UCIDC10TC10.txt

Accuracy = 97.4215% (2607/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC20.txt BACMQ100UCIDC10.txt.model testQ100UCIDC10TC20.txt

Accuracy = 98.3931% (2633/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC30.txt BACMQ100UCIDC10.txt.model testQ100UCIDC10TC30.txt

Accuracy = 98.9163% (2647/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC50.txt BACMQ100UCIDC10.txt.model testQ100UCIDC10TC50.txt

Accuracy = 99.0284% (2650/2676) (classification)




% cross validation test that tested by BACMQ100UCIDC20.txt.model for all test dataset

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC01.txt BACMQ100UCIDC20.txt.model testQ100UCIDC20TC01.txt

Accuracy = 68.3109% (1828/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC02.txt BACMQ100UCIDC20.txt.model testQ100UCIDC20TC02.txt

Accuracy = 72.6831% (1945/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC05.txt BACMQ100UCIDC20.txt.model testQ100UCIDC20TC05.txt

Accuracy = 84.6786% (2266/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC10.txt BACMQ100UCIDC20.txt.model testQ100UCIDC20TC10.txt

Accuracy = 93.2362% (2495/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC20.txt BACMQ100UCIDC20.txt.model testQ100UCIDC20TC20.txt

Accuracy = 98.8042% (2644/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC30.txt BACMQ100UCIDC20.txt.model testQ100UCIDC20TC30.txt

Accuracy = 99.1031% (2652/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC50.txt BACMQ100UCIDC20.txt.model testQ100UCIDC20TC50.txt

Accuracy = 99.29% (2657/2676) (classification)




% cross validation test that tested by BACMQ100UCIDC30.txt.model for all test dataset

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC01.txt BACMQ100UCIDC30.txt.model testQ100UCIDC30TC01.txt

Accuracy = 58.4081% (1563/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC02.txt BACMQ100UCIDC30.txt.model testQ100UCIDC30TC02.txt

Accuracy = 63.154% (1690/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC05.txt BACMQ100UCIDC30.txt.model testQ100UCIDC30TC05.txt

Accuracy = 74.6263% (1997/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC10.txt BACMQ100UCIDC30.txt.model testQ100UCIDC30TC10.txt

Accuracy = 86.0987% (2304/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC20.txt BACMQ100UCIDC30.txt.model testQ100UCIDC30TC20.txt

Accuracy = 95.3662% (2552/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC30.txt BACMQ100UCIDC30.txt.model testQ100UCIDC30TC30.txt

Accuracy = 99.2526% (2656/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC50.txt BACMQ100UCIDC30.txt.model testQ100UCIDC30TC50.txt

Accuracy = 99.4395% (2661/2676) (classification)




% cross validation test that tested by BACMQ100UCIDC50.txt.model for all test dataset

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC01.txt BACMQ100UCIDC50.txt.model testQ100UCIDC50TC01.txt

Accuracy = 52.0553% (1393/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC02.txt BACMQ100UCIDC50.txt.model testQ100UCIDC50TC02.txt

Accuracy = 53.0269% (1419/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC05.txt BACMQ100UCIDC50.txt.model testQ100UCIDC50TC05.txt

Accuracy = 58.1465% (1556/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC10.txt BACMQ100UCIDC50.txt.model testQ100UCIDC50TC10.txt

Accuracy = 68.2362% (1826/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC20.txt BACMQ100UCIDC50.txt.model testQ100UCIDC50TC20.txt

Accuracy = 84.716% (2267/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC30.txt BACMQ100UCIDC50.txt.model testQ100UCIDC50TC30.txt

Accuracy = 94.6188% (2532/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC50.txt BACMQ100UCIDC50.txt.model testQ100UCIDC50TC50.txt

Accuracy = 99.6263% (2666/2676) (classification)

% cross validation test, tested by 7 models, including the mixed set itself.

Phoenix:windows Phoenix$ svm-predict BACMQ100Mixed7C.txt BACMQ100UCIDC01.txt.model testQ100UCIDC01TCMixed7.txt

Accuracy = 95.7399% (2562/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100Mixed7C.txt BACMQ100UCIDC02.txt.model testQ100UCIDC02TCMixed7.txt

Accuracy = 95.5157% (2556/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100Mixed7C.txt BACMQ100UCIDC05.txt.model testQ100UCIDC05TCMixed7.txt

Accuracy = 95.2541% (2549/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100Mixed7C.txt BACMQ100UCIDC10.txt.model testQ100UCIDC10TCMixed7.txt

Accuracy = 92.3767% (2472/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100Mixed7C.txt BACMQ100UCIDC20.txt.model testQ100UCIDC20TCMixed7.txt

Accuracy = 86.9955% (2328/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100Mixed7C.txt BACMQ100UCIDC30.txt.model testQ100UCIDC30TCMixed7.txt

Accuracy = 81.6891% (2186/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100Mixed7C.txt BACMQ100UCIDC50.txt.model testQ100UCIDC50TCMixed7.txt

Accuracy = 73.7294% (1973/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100Mixed7C.txt BACMQ100Mixed7C.txt.model testQ100UCIDCMixed7TCMixed7.txt

Accuracy = 96.0762% (2571/2676) (classification)

% cross validation test, tested by the mixed set.




Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC01.txt BACMQ100Mixed7C.txt.model testQ100UCIDCMixed7TC01.txt

Accuracy = 88.0045% (2355/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC02.txt BACMQ100Mixed7C.txt.model testQ100UCIDCMixed7TC02.txt

Accuracy = 90.1719% (2413/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC05.txt BACMQ100Mixed7C.txt.model testQ100UCIDCMixed7TC05.txt

Accuracy = 94.4694% (2528/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC10.txt BACMQ100Mixed7C.txt.model testQ100UCIDCMixed7TC10.txt

Accuracy = 96.9731% (2595/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC20.txt BACMQ100Mixed7C.txt.model testQ100UCIDCMixed7TC20.txt

Accuracy = 98.5426% (2637/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC30.txt BACMQ100Mixed7C.txt.model testQ100UCIDCMixed7TC30.txt

Accuracy = 98.9537% (2648/2676) (classification)

Phoenix:windows Phoenix$ svm-predict BACMQ100UCIDC50.txt BACMQ100Mixed7C.txt.model testQ100UCIDCMixed7TC50.txt

Accuracy = 99.0658% (2651/2676) (classification)

8. Report the tamper detection accuracy for seam carving detection for compressed images at QF100 as in Table II.