IsmailPeters - Revision history

imported>Wandell: /* Code */

2013-03-21T22:49:26Z

Code

← Older revision		Revision as of 22:49, 21 March 2013
Line 134:		Line 134:

	==Code==		==Code==
	[[File:IsmailPeters_CameraForensics.zip]]		[[File:IsmailPeters_CameraForensics.zip Source Code]]

	= Appendix II - Work partition =		= Appendix II - Work partition =
	Partners through and through. 50/50.<br>		Partners through and through. 50/50.<br>
	Although Chasen developed all of the sweet UI tests.		Although Chasen developed all of the sweet UI tests.

imported>Projects221: /* Final Test Images */

2013-03-20T18:59:56Z

Final Test Images

← Older revision		Revision as of 18:59, 20 March 2013
Line 117:		Line 117:
	[[File:jpeg_fake.png\|thumb\|300px\|center\|Figure 18: Similarity plot for tampered jpg image]]		[[File:jpeg_fake.png\|thumb\|300px\|center\|Figure 18: Similarity plot for tampered jpg image]]
	[[File:norm_fake.png\|thumb\|300px\|center\|Figure 17: Similarity plot for tampered png image]]		[[File:norm_fake.png\|thumb\|300px\|center\|Figure 17: Similarity plot for tampered png image]]
	[[File:resize_fake.png\|thumb\|300px\|center\|Figure 18: Similarity plot for ~~tampered jpg~~ image]]		[[File:resize_fake.png\|thumb\|300px\|center\|Figure 18: Similarity plot for re-sized png image]]

	= Conclusions =		= Conclusions =

imported>Projects221: /* Final Test Images */

2013-03-20T18:57:17Z

Final Test Images

← Older revision		Revision as of 18:57, 20 March 2013
Line 112:		Line 112:

	==Final Test Images ==		==Final Test Images ==
	Of the final 11 sets of test images that were provided, the classifier was able to correctly categorize 5 properly(fake classified as fake, real classified as real). The algorithm is still highly sensitive to re-sizing of the entire image and JPEG compression after the image has been modified. Figures X,Y demonstrate the issues seen with jpg compression and why it is difficult to determine a single threshold value.		Of the final 11 sets of test images that were provided, the classifier was able to correctly categorize 5 properly(fake classified as fake, real classified as real). The algorithm is still highly sensitive to re-sizing of the entire image and JPEG compression after the image has been modified. Figures X,Y demonstrate the issues seen with jpg compression and why it is difficult to determine a single threshold value for both images. Similarly, figures 19 and 20 demonstrate the issues seen by re sizing the tampered image.

	[[File:png_fake.png\|thumb\|300px\|center\|Figure 17: Similarity plot for tampered png image]]		[[File:png_fake.png\|thumb\|300px\|center\|Figure 17: Similarity plot for tampered png image]]
	[[File:jpeg_fake.png\|thumb\|300px\|center\|Figure 18: Similarity plot for tampered jpg image]]		[[File:jpeg_fake.png\|thumb\|300px\|center\|Figure 18: Similarity plot for tampered jpg image]]
			[[File:norm_fake.png\|thumb\|300px\|center\|Figure 17: Similarity plot for tampered png image]]
			[[File:resize_fake.png\|thumb\|300px\|center\|Figure 18: Similarity plot for tampered jpg image]]

	= Conclusions =		= Conclusions =

imported>Projects221: /* Expectation/Maximization */

2013-03-20T18:55:12Z

Expectation/Maximization

← Older revision		Revision as of 18:55, 20 March 2013
Line 40:		Line 40:
	:<math>f(x,y) = \sum_{u,v=-N}^{N} \alpha_{u,v}f(x+u,y+v)+n(x,y)</math>		:<math>f(x,y) = \sum_{u,v=-N}^{N} \alpha_{u,v}f(x+u,y+v)+n(x,y)</math>

	where ''<math>\alpha_{u,v}</math>'' are the interpolation coefficients and ''n(x,y)'' is the residual error of the linear model and ideally a Gaussian distribution with zero mean. The probability ~~of observing~~ a sample ''f(x,y)'' ~~given~~ that it is correlated to its neighbors is given by the equation:		where ''<math>\alpha_{u,v}</math>'' are the interpolation coefficients and ''n(x,y)'' is the residual error of the linear model and ideally a Gaussian distribution with zero mean. The probability that a sample ''f(x,y)'' exists and that it is correlated to its neighbors is given by the equation:

	:<math>P(x,y) = \frac{1}{\sigma \sqrt{2\pi}}exp\left[-\frac{1}{2\sigma^{2}}\left(f(x,y)-\sum_{u,v=-N}^{N} \alpha_{u,v}f(x+u,y+v)\right)^2\right]</math>		:<math>P(x,y) = \frac{1}{\sigma \sqrt{2\pi}}exp\left[-\frac{1}{2\sigma^{2}}\left(f(x,y)-\sum_{u,v=-N}^{N} \alpha_{u,v}f(x+u,y+v)\right)^2\right]</math>

imported>Projects221: /* Final Test Images */

2013-03-20T18:54:32Z

Final Test Images

← Older revision		Revision as of 18:54, 20 March 2013
Line 113:		Line 113:
	==Final Test Images ==		==Final Test Images ==
	Of the final 11 sets of test images that were provided, the classifier was able to correctly categorize 5 properly(fake classified as fake, real classified as real). The algorithm is still highly sensitive to re-sizing of the entire image and JPEG compression after the image has been modified. Figures X,Y demonstrate the issues seen with jpg compression and why it is difficult to determine a single threshold value.		Of the final 11 sets of test images that were provided, the classifier was able to correctly categorize 5 properly(fake classified as fake, real classified as real). The algorithm is still highly sensitive to re-sizing of the entire image and JPEG compression after the image has been modified. Figures X,Y demonstrate the issues seen with jpg compression and why it is difficult to determine a single threshold value.

			[[File:png_fake.png\|thumb\|300px\|center\|Figure 17: Similarity plot for tampered png image]]
			[[File:jpeg_fake.png\|thumb\|300px\|center\|Figure 18: Similarity plot for tampered jpg image]]

	= Conclusions =		= Conclusions =

imported>Projects221: /* Expectation/Maximization */

2013-03-20T18:54:31Z

Expectation/Maximization

← Older revision		Revision as of 18:54, 20 March 2013
Line 40:		Line 40:
	:<math>f(x,y) = \sum_{u,v=-N}^{N} \alpha_{u,v}f(x+u,y+v)+n(x,y)</math>		:<math>f(x,y) = \sum_{u,v=-N}^{N} \alpha_{u,v}f(x+u,y+v)+n(x,y)</math>

	where ''<math>\alpha_{u,v}</math>'' are the interpolation coefficients and ''n(x,y)'' is the residual error of the linear model and ideally a Gaussian distribution with zero mean. The probability ~~that~~ a sample ''f(x,y)'' ~~exists~~ given that it is correlated to its neighbors is given by the equation:		where ''<math>\alpha_{u,v}</math>'' are the interpolation coefficients and ''n(x,y)'' is the residual error of the linear model and ideally a Gaussian distribution with zero mean. The probability of observing a sample ''f(x,y)'' given that it is correlated to its neighbors is given by the equation:

	:<math>P(x,y) = \frac{1}{\sigma \sqrt{2\pi}}exp\left[-\frac{1}{2\sigma^{2}}\left(f(x,y)-\sum_{u,v=-N}^{N} \alpha_{u,v}f(x+u,y+v)\right)^2\right]</math>		:<math>P(x,y) = \frac{1}{\sigma \sqrt{2\pi}}exp\left[-\frac{1}{2\sigma^{2}}\left(f(x,y)-\sum_{u,v=-N}^{N} \alpha_{u,v}f(x+u,y+v)\right)^2\right]</math>
Line 67:		Line 67:

	In this way, a new set of coefficients can then be easily computed. This rearrangement of the data provided much faster computations, and it was very easy to add more coefficients to the system. The EM algorithm was allowed to run until the coefficients generated in successive iterations differed by an amount less than some arbitrarily small threshold.		In this way, a new set of coefficients can then be easily computed. This rearrangement of the data provided much faster computations, and it was very easy to add more coefficients to the system. The EM algorithm was allowed to run until the coefficients generated in successive iterations differed by an amount less than some arbitrarily small threshold.

	=== 2-D FFT ===		=== 2-D FFT ===
	Once a probability map was generated from the EM algorithm, the next step was to determine if the periodic pattern of the CFA was present in the map. To do this, a two-dimensional FFT was performed on the probability map. The default 2-D FFT magnitude plot in MATLAB contains all of the low frequency information in the upper left corner, and the high frequency components along the left and bottom edges. The plot also scales to the largest value. In order to make the desired periodicity easier to observe, the FFT plot was up-sampled, high-pass filtered, and shifted so that the low frequency components are at the center of the image. Figure 2 shows the original FFT plot for a given probability map. Figure 3 is the same data, but it has been upsampled and filtered in order to bring forward the periodic effects of the CFA interpolation.		Once a probability map was generated from the EM algorithm, the next step was to determine if the periodic pattern of the CFA was present in the map. To do this, a two-dimensional FFT was performed on the probability map. The default 2-D FFT magnitude plot in MATLAB contains all of the low frequency information in the upper left corner, and the high frequency components along the left and bottom edges. The plot also scales to the largest value. In order to make the desired periodicity easier to observe, the FFT plot was up-sampled, high-pass filtered, and shifted so that the low frequency components are at the center of the image. Figure 2 shows the original FFT plot for a given probability map. Figure 3 is the same data, but it has been upsampled and filtered in order to bring forward the periodic effects of the CFA interpolation.

imported>Projects221: /* Final Test Images */

2013-03-20T18:52:35Z

Final Test Images

← Older revision		Revision as of 18:52, 20 March 2013
Line 111:		Line 111:

	==Final Test Images ==		==Final Test Images ==
	Of the final 11 sets of test images that were provided, the classifier was able to correctly categorize 5 properly(fake classified as fake, real classified as real). ~~As seen below,~~ The algorithm is ~~very~~ sensitive to ~~image~~ re-sizing and JPEG compression.		Of the final 11 sets of test images that were provided, the classifier was able to correctly categorize 5 properly(fake classified as fake, real classified as real). The algorithm is still highly sensitive to re-sizing of the entire image and JPEG compression after the image has been modified. Figures X,Y demonstrate the issues seen with jpg compression and why it is difficult to determine a single threshold value.

	= Conclusions =		= Conclusions =

imported>Projects221 at 18:46, 20 March 2013

2013-03-20T18:46:46Z

← Older revision		Revision as of 18:46, 20 March 2013
Line 110:		Line 110:
	[[File:similarity.gif\|thumb\|300px\|center\|Figure 16: Corresponding similarity plot]]		[[File:similarity.gif\|thumb\|300px\|center\|Figure 16: Corresponding similarity plot]]

	~~<br>~~		==Final Test Images ==
			Of the final 11 sets of test images that were provided, the classifier was able to correctly categorize 5 properly(fake classified as fake, real classified as real). As seen below, The algorithm is very sensitive to image re-sizing and JPEG compression.

	= Conclusions =		= Conclusions =
	The algorithms provided could still use further improvements as they stand. First, the EM algorithm has trouble working with image windows of constant color. This low spatial frequency content make it hard for the algorithm to separate the pixels correlated with their neighbors and the ones that are independent. If this detail can be teased out of the algorithm, the similarity plots would be much more stable for many more of the images. Another improvement that can be made is figuring out how to increase the accuracy of the classifier. As of right now, the classifier is very sensitive to ~~resizing~~ or ~~jpg~~ compression of the image after it has been modified. Possible methods to improve the hit rate would be decreasing the image window and step size. One possible reason that the accuracy is reduced is that the sweep window is too large and not sensitive to tampering.		The algorithms provided could still use further improvements as they stand. First, the EM algorithm has trouble working with image windows of constant color. This low spatial frequency content make it hard for the algorithm to separate the pixels correlated with their neighbors and the ones that are independent. If this detail can be teased out of the algorithm, the similarity plots would be much more stable for many more of the images. Another improvement that can be made is figuring out how to increase the accuracy of the classifier. As of right now, the classifier is very sensitive to re-sizing or JPEG compression of the image after it has been modified. Possible methods to improve the hit rate would be decreasing the image window and step size. One possible reason that the accuracy is reduced is that the sweep window is too large and not sensitive to tampering.
	<br>		<br>
	As it stands, the classifier can successfully identify a fake image as fake and real image as real for 41 test pairs of the original 70 test pairs given.		As it stands, the classifier can successfully identify a fake image as fake and real image as real for 41 test pairs of the original 70 test pairs given.

imported>Projects221: /* Conclusions */

2013-03-20T18:42:25Z

Conclusions

← Older revision		Revision as of 18:42, 20 March 2013
Line 113:		Line 113:

	= Conclusions =		= Conclusions =
	The algorithms provided could still use further improvements as they stand. First, the EM algorithm has trouble working with image windows of constant color. This low spatial frequency content make it hard for the algorithm to separate the pixels correlated with their neighbors and the ones that are independent. If this detail can be teased out of the algorithm, the similarity plots would be much more stable for many more of the images. Another improvement that can be made is figuring out how to increase the accuracy of the classifier. Possible methods to improve the hit rate would be decreasing the image window and step size. One possible reason that the accuracy is reduced is that the sweep window is too large and not sensitive to tampering.		The algorithms provided could still use further improvements as they stand. First, the EM algorithm has trouble working with image windows of constant color. This low spatial frequency content make it hard for the algorithm to separate the pixels correlated with their neighbors and the ones that are independent. If this detail can be teased out of the algorithm, the similarity plots would be much more stable for many more of the images. Another improvement that can be made is figuring out how to increase the accuracy of the classifier. As of right now, the classifier is very sensitive to resizing or jpg compression of the image after it has been modified. Possible methods to improve the hit rate would be decreasing the image window and step size. One possible reason that the accuracy is reduced is that the sweep window is too large and not sensitive to tampering.
	<br>		<br>
	As it stands, the classifier can successfully identify a fake image as fake and real image as real for 41 test pairs of the original 70 test pairs given.		As it stands, the classifier can successfully identify a fake image as fake and real image as real for 41 test pairs of the original 70 test pairs given.

imported>Projects221: /* Conclusions */

2013-03-20T17:05:52Z

Conclusions

← Older revision		Revision as of 17:05, 20 March 2013
Line 115:		Line 115:
	The algorithms provided could still use further improvements as they stand. First, the EM algorithm has trouble working with image windows of constant color. This low spatial frequency content make it hard for the algorithm to separate the pixels correlated with their neighbors and the ones that are independent. If this detail can be teased out of the algorithm, the similarity plots would be much more stable for many more of the images. Another improvement that can be made is figuring out how to increase the accuracy of the classifier. Possible methods to improve the hit rate would be decreasing the image window and step size. One possible reason that the accuracy is reduced is that the sweep window is too large and not sensitive to tampering.		The algorithms provided could still use further improvements as they stand. First, the EM algorithm has trouble working with image windows of constant color. This low spatial frequency content make it hard for the algorithm to separate the pixels correlated with their neighbors and the ones that are independent. If this detail can be teased out of the algorithm, the similarity plots would be much more stable for many more of the images. Another improvement that can be made is figuring out how to increase the accuracy of the classifier. Possible methods to improve the hit rate would be decreasing the image window and step size. One possible reason that the accuracy is reduced is that the sweep window is too large and not sensitive to tampering.
	<br>		<br>
	As it stands, the classifier can successfully identify a fake image as fake and real image as real for 41 test pairs.		As it stands, the classifier can successfully identify a fake image as fake and real image as real for 41 test pairs of the original 70 test pairs given.

	= References =		= References =
	1. Popescu, A.C.; Farid, H.,''Exposing digital forgeries in color filter array interpolated images'', Signal Processing, IEEE Transactions on , vol.53, no.10, pp.3948,3959, Oct. 2005		1. Popescu, A.C.; Farid, H.,''Exposing digital forgeries in color filter array interpolated images'', Signal Processing, IEEE Transactions on , vol.53, no.10, pp.3948,3959, Oct. 2005