SchmittWerner - Revision history

imported>Projects221: /* Future Work */

2013-03-20T16:38:58Z

Future Work

← Older revision		Revision as of 16:38, 20 March 2013
Line 137:		Line 137:
	Future research should focus on abstracting the metric thresholding method developed in this paper to a higher dimensional space. It was experimentally determined that additional reasonably orthogonal high-performing metrics allowed SVM to perform better than our proposed approach. However, the lack of direct tuning of the false positive rate kept this method from being a useful technique. If the false positive rate could be controlled in an SVM setting, it is theorized that the classification results would improve considerably.		Future research should focus on abstracting the metric thresholding method developed in this paper to a higher dimensional space. It was experimentally determined that additional reasonably orthogonal high-performing metrics allowed SVM to perform better than our proposed approach. However, the lack of direct tuning of the false positive rate kept this method from being a useful technique. If the false positive rate could be controlled in an SVM setting, it is theorized that the classification results would improve considerably.

	Also, to improve on the thresholding mechanism adopted in this paper, future work should consider constructing a Receiver Operating Curve (ROC) to determine the optimal classification parameters. Various metrics and their corresponding thresholds should be used on the test data. The metric and the threshold value that create the largest area under the ~~operating curve~~ should be chosen as the optimal classification parameters.		Also, to improve on the thresholding mechanism adopted in this paper, future work should consider constructing a Receiver Operating Curve (ROC) to determine the optimal classification parameters. Various metrics and their corresponding thresholds should be used on the test data. The metric and the threshold value that create the largest area under the ROC should be chosen as the optimal classification parameters.

	= References - Resources and related work =		= References - Resources and related work =

imported>Projects221: /* Future Work */

2013-03-20T16:38:39Z

Future Work

← Older revision		Revision as of 16:38, 20 March 2013
Line 137:		Line 137:
	Future research should focus on abstracting the metric thresholding method developed in this paper to a higher dimensional space. It was experimentally determined that additional reasonably orthogonal high-performing metrics allowed SVM to perform better than our proposed approach. However, the lack of direct tuning of the false positive rate kept this method from being a useful technique. If the false positive rate could be controlled in an SVM setting, it is theorized that the classification results would improve considerably.		Future research should focus on abstracting the metric thresholding method developed in this paper to a higher dimensional space. It was experimentally determined that additional reasonably orthogonal high-performing metrics allowed SVM to perform better than our proposed approach. However, the lack of direct tuning of the false positive rate kept this method from being a useful technique. If the false positive rate could be controlled in an SVM setting, it is theorized that the classification results would improve considerably.

	Also, to improve on the thresholding mechanism adopted in this paper, future work should consider constructing a Receiver Operating Curve (ROC) ~~for various values of~~ the ~~threshold~~. The ~~value of~~ the threshold that ~~places~~ the operating ~~point at the optimal point on the~~ curve should be chosen as the classification ~~threshold~~.		Also, to improve on the thresholding mechanism adopted in this paper, future work should consider constructing a Receiver Operating Curve (ROC) to determine the optimal classification parameters. Various metrics and their corresponding thresholds should be used on the test data. The metric and the threshold value that create the largest area under the operating curve should be chosen as the optimal classification parameters.

	= References - Resources and related work =		= References - Resources and related work =

imported>Projects221: /* Future Work */

2013-03-20T16:35:29Z

Future Work

← Older revision		Revision as of 16:35, 20 March 2013
Line 136:		Line 136:

	Future research should focus on abstracting the metric thresholding method developed in this paper to a higher dimensional space. It was experimentally determined that additional reasonably orthogonal high-performing metrics allowed SVM to perform better than our proposed approach. However, the lack of direct tuning of the false positive rate kept this method from being a useful technique. If the false positive rate could be controlled in an SVM setting, it is theorized that the classification results would improve considerably.		Future research should focus on abstracting the metric thresholding method developed in this paper to a higher dimensional space. It was experimentally determined that additional reasonably orthogonal high-performing metrics allowed SVM to perform better than our proposed approach. However, the lack of direct tuning of the false positive rate kept this method from being a useful technique. If the false positive rate could be controlled in an SVM setting, it is theorized that the classification results would improve considerably.

			Also, to improve on the thresholding mechanism adopted in this paper, future work should consider constructing a Receiver Operating Curve (ROC) for various values of the threshold. The value of the threshold that places the operating point at the optimal point on the curve should be chosen as the classification threshold.

	= References - Resources and related work =		= References - Resources and related work =

imported>Projects221: /* Results */

2013-03-18T22:20:31Z

Results

← Older revision		Revision as of 22:20, 18 March 2013
Line 123:		Line 123:
	[[File:Figure10_sw.jpg\|thumb\|center\|900px\|Figure 12. Undetected small tampered region]] [[File:Figure11_sw.jpg\|thumb\|center\|900px\|Figure 13. Untampered counterpart image ]]		[[File:Figure10_sw.jpg\|thumb\|center\|900px\|Figure 12. Undetected small tampered region]] [[File:Figure11_sw.jpg\|thumb\|center\|900px\|Figure 13. Untampered counterpart image ]]

	After running our proposed algorithm on the supplied test images given after the presentation, there are a few key results we should comment on. First, our algorithm appears to suffer when applied to a JPEG compressed image. The artifacts introduced by the quantization step and the macro-blocking step of the compression procedure greatly modify the periodicities introduced in the CFA demosaicing algorithm.		After running our proposed algorithm on the supplied test images given after the presentation, there are a few key results we should comment on. First, our algorithm appears to suffer when applied to a JPEG compressed image. The artifacts introduced by the quantization step and the macro-blocking step of the compression procedure greatly modify the periodicities introduced in the CFA demosaicing algorithm. Furthermore, one can view the JPEG quantization step as a lowpass filtering in the frequency domain. Because our algorithm looks for magnitude spectrum peaks at certain locations, lowpass filtering directly effects the efficacy of our peak-finding metrics.

	Furthermore, many of the images in the test suite had modifications near the borders of the images. To avoid bordering artifacts introduced by edges, our algorithm does not train on bordering patches. Thus, tampering regions on the borders of images are difficult for our algorithm to recognize. Finally, as we mentioned earlier, tampered regions that are smaller than window_size by window_size are challenging for our algorithm to identify. A modification to take into account smaller tampered regions would be to employ a multi-scale approach on top of our algorithm.		Furthermore, many of the images in the test suite had modifications near the borders of the images. To avoid bordering artifacts introduced by edges, our algorithm does not train on bordering patches. Thus, tampering regions on the borders of images are difficult for our algorithm to recognize. Finally, as we mentioned earlier, tampered regions that are smaller than window_size by window_size are challenging for our algorithm to identify. A modification to take into account smaller tampered regions would be to employ a multi-scale approach on top of our algorithm.

imported>Projects221: /* Results */

2013-03-18T22:18:13Z

Results

← Older revision		Revision as of 22:18, 18 March 2013
Line 122:		Line 122:

	[[File:Figure10_sw.jpg\|thumb\|center\|900px\|Figure 12. Undetected small tampered region]] [[File:Figure11_sw.jpg\|thumb\|center\|900px\|Figure 13. Untampered counterpart image ]]		[[File:Figure10_sw.jpg\|thumb\|center\|900px\|Figure 12. Undetected small tampered region]] [[File:Figure11_sw.jpg\|thumb\|center\|900px\|Figure 13. Untampered counterpart image ]]

			After running our proposed algorithm on the supplied test images given after the presentation, there are a few key results we should comment on. First, our algorithm appears to suffer when applied to a JPEG compressed image. The artifacts introduced by the quantization step and the macro-blocking step of the compression procedure greatly modify the periodicities introduced in the CFA demosaicing algorithm.

			Furthermore, many of the images in the test suite had modifications near the borders of the images. To avoid bordering artifacts introduced by edges, our algorithm does not train on bordering patches. Thus, tampering regions on the borders of images are difficult for our algorithm to recognize. Finally, as we mentioned earlier, tampered regions that are smaller than window_size by window_size are challenging for our algorithm to identify. A modification to take into account smaller tampered regions would be to employ a multi-scale approach on top of our algorithm.

	= Conclusions =		= Conclusions =

imported>Projects221: /* Results */

2013-03-18T22:02:21Z

Results

← Older revision		Revision as of 22:02, 18 March 2013
Line 98:		Line 98:
	= Results =		= Results =

	We use a image data set provided by the authors of [3] to train and test the algorithm described in this paper. The results below correspond to training the thresholding algorithm on 90% of the data set. Then the trained algorithm is queried using the remaining 10% of the data set. An example ~~"fake" and its counterpart "~~real~~" image are shown in Figures 8~~ and ~~9. The modifications in these~~ images are too small to be apparent to the human eye. ~~Figure 10 shows the~~ pixel level differences ~~between Figures 8 and 9 (~~increased by a gamma factor of 8 for visibility).		We use a image data set provided by the authors of [3] to train and test the algorithm described in this paper. The results below correspond to training the thresholding algorithm on 90% of the data set. Then the trained algorithm is queried using the remaining 10% of the data set. The modifications between example real and fake images are too small to be apparent to the human eye and are shown in Figure 8. These pixel level differences are increased by a gamma factor of 8 for visibility.

	~~[[File:FakeImage_sw.jpg\|thumb\|center\|900px\|Figure 8. Tampered image ]]~~		[[File:DiffImage_sw.jpg\|thumb\|center\|900px\|Figure 10. Raw pixel differences between real and fake images increased by a gamma factor of 8 for visibility ]]
	~~[[File:RealImage_sw.jpg\|thumb\|center\|900px\|Figure 9. Untampered counterpart image ]]~~
	[[File:DiffImage_sw.jpg\|thumb\|center\|900px\|Figure 10. Raw pixel differences between ~~Figures 8~~ and ~~9 (~~increased by a gamma factor of 8 for visibility) ]]

	The efficacy of the method discussed in this paper can be visualized using a confusion matrix. Such a matrix indicates the number of true positives, false positives, true negatives, and false negatives in a tabular format. We define a true positive to be an actually tampered image declared as a tampered image and a false negative as an actually tampered image declared as an untampered image, etc. Figure 11 shows the confusion matrix computed from the testing results.		The efficacy of the method discussed in this paper can be visualized using a confusion matrix. Such a matrix indicates the number of true positives, false positives, true negatives, and false negatives in a tabular format. We define a true positive to be an actually tampered image declared as a tampered image and a false negative as an actually tampered image declared as an untampered image, etc. Figure 9 shows the confusion matrix computed from the testing results.

	:{\| border="1"		:{\| border="1"
Line 112:		Line 110:
	\| False Positives <br/> 2/20		\| False Positives <br/> 2/20
	\| True Negatives <br/> 8/20		\| True Negatives <br/> 8/20
	\|+ Figure 11. The confusion matrix		\|+ Figure 9. The confusion matrix
	\|}		\|}

	The reader should note that both the false positive and false negative rates are smaller than the true positive and true negative rates, which is an indication of acceptable performance of the algorithm. Furthermore, the false positive limitation requirement manifests itself in a very small value in the confusion matrix. We should point out that the false positive rate requirement was set per patch, not for the overall image. And an image is declared as false if just one of its patches is declared as fake. Therefore, the 10% false positive rate shown in the confusion matrix does not contradict the 1% false positive condition posited in the thresholding section. Because there was no limitation placed on the false negative rate, this score is acceptably approximately twice as high.		The reader should note that both the false positive and false negative rates are smaller than the true positive and true negative rates, which is an indication of acceptable performance of the algorithm. Furthermore, the false positive limitation requirement manifests itself in a very small value in the confusion matrix. We should point out that the false positive rate requirement was set per patch, not for the overall image. And an image is declared as false if just one of its patches is declared as fake. Therefore, the 10% false positive rate shown in the confusion matrix does not contradict the 1% false positive condition posited in the thresholding section. Because there was no limitation placed on the false negative rate, this score is acceptably approximately twice as high.

	Our algorithm performs well when the tampered region has dimensions of approximately window_size by window_size. In general, the smaller the tampered region is in comparison to the raster scanning window, the worse the algorithm performs. Figures 12 and 13 show a case where our algorithm correctly identifies a tampered region in an image, and correctly neglects to identify any tampered regions in the untampered counterpart image. Notice the similarity between Figures 8 and 12. These two figures derive from the same tampered image and so it is expected that the detected tampered region in Figure 12 spatially lies within the green rectangle of Figure 8.		Our algorithm performs well when the tampered region has dimensions of approximately window_size by window_size. In general, the smaller the tampered region is in comparison to the raster scanning window, the worse the algorithm performs. Figures 10 and 11 show a case where our algorithm correctly identifies a tampered region in an image, and correctly neglects to identify any tampered regions in the untampered counterpart image. Notice the similarity between Figures 8 and 10. These two figures derive from the same tampered image and so it is expected that the detected tampered region in Figure 10 spatially lies within the green rectangle of Figure 8.

	Figures 14 and 15, on the other hand, show a case where a small tampered region cannot be detected by our algorithm. In these figures, the actual tampered regions are denoted by a green rectangle, and our algorithms tampering detection is denoted by black regions.		Figures 12 and 13, on the other hand, show a case where a small tampered region cannot be detected by our algorithm. In these figures, the actual tampered regions are denoted by a green rectangle, and our algorithms tampering detection is denoted by black regions.

	[[File:Figure8_sw.jpg\|thumb\|center\|900px\|Figure 12. Correctly identified tampered region]] [[File:Figure9_sw.jpg\|thumb\|center\|900px\|Figure 13. Untampered counterpart image ]]		[[File:Figure8_sw.jpg\|thumb\|center\|900px\|Figure 10. Correctly identified tampered region]] [[File:Figure9_sw.jpg\|thumb\|center\|900px\|Figure 11. Untampered counterpart image ]]

	[[File:Figure10_sw.jpg\|thumb\|center\|900px\|Figure 14. Undetected small tampered region]] [[File:Figure11_sw.jpg\|thumb\|center\|900px\|Figure 15. Untampered counterpart image ]]		[[File:Figure10_sw.jpg\|thumb\|center\|900px\|Figure 12. Undetected small tampered region]] [[File:Figure11_sw.jpg\|thumb\|center\|900px\|Figure 13. Untampered counterpart image ]]

	= Conclusions =		= Conclusions =

imported>Projects221: /* Introduction */

2013-03-18T21:00:24Z

Introduction

← Older revision		Revision as of 21:00, 18 March 2013
Line 10:		Line 10:
	Due to improvements in digital technologies, image forgeries are increasingly believable and easy to produce. There are some situations, including forensic settings, where it is necessary to detect and localize such manipulations in images. A robust technique to identify image forgeries on uncompressed images is presented in this paper.		Due to improvements in digital technologies, image forgeries are increasingly believable and easy to produce. There are some situations, including forensic settings, where it is necessary to detect and localize such manipulations in images. A robust technique to identify image forgeries on uncompressed images is presented in this paper.

	Cameras introduce specific statistical correlations in images [1]. One such identifiable pattern is caused by color filter arrays (CFA). Typically, each pixel sensor only measures the value corresponding to one color channel. The most common mosaic arrangement is called the Bayer Filter, which is shown in Figure 1. To produce red, green, and blue (R,G,B) values at each pixel, demosaicing algorithms are used to interpolate missing color channel values by taking a combination of neighbor’s color channels [2]. The authors of [3] suggest that it is sufficient to treat the color channels independently and assume a linear combination is used for interpolation. While we assume the linear model in this paper, we show that considering the dependence between separate color channels produces more accurate results.		Cameras introduce specific statistical correlations in images [1]. One such identifiable pattern is caused by color filter arrays (CFA). Typically, each pixel sensor only measures the value corresponding to one color channel. The most common mosaic arrangement is called the Bayer Filter, which is shown in Figure 1. To produce red, green, and blue (R,G,B) values at each pixel, demosaicing algorithms are used to interpolate missing color channel values by taking a combination of neighbor’s color channels [2]. The authors of [3] suggest that it is sufficient to treat the color channels independently and assume that a linear combination is used for interpolation. While we follow this advice and assume the linear model in this paper, we show that considering the dependence between separate color channels produces more accurate results.

	[[File:Figure1_sw.jpg\|thumb\|center\|900px\|Figure 1. Color Filter Array technology [3]]]		[[File:Figure1_sw.jpg\|thumb\|center\|900px\|Figure 1. Color Filter Array technology [5]]]

	In the seminal paper by Farid [4], the assumption of linearity allows an expectation maximization algorithm (EM) to be used. This predicts whether each pixel is likely to be a linear combination of its neighboring pixels in two-dimensions. Periodicities in this likelihood map appear as strong peaks in the Fourier magnitude spectrum. In regions where tampering is present, these peaks will appear in different locations, or not at all. A follow-up paper [3] expands this technique to look for the periodicities introduced by CFA demosaicing algorithms in particular. In regions where tampering is present, the correlations introduced by CFA are obscured. Our paper attempts to reproduce the results found in [3] and provide additional refinements to improve the efficacy of forgery detection.		In the seminal paper by Farid [4], the assumption of linearity allows an expectation maximization algorithm (EM) to be used. This predicts whether each pixel is likely to be a linear combination of its neighboring pixels in two-dimensions. Periodicities in this likelihood map appear as strong peaks in the Fourier magnitude spectrum. In regions where tampering is present, these peaks will appear in different locations, or not at all. A follow-up paper [3] expands this technique to look for the periodicities introduced by CFA demosaicing algorithms in particular. In regions where tampering is present, the correlations introduced by CFA are obscured. Our paper attempts to reproduce the results found in [3] and provide additional refinements to improve the efficacy of forgery detection.

	= Methods =		= Methods =

imported>Projects221: /* Code */

2013-03-18T20:58:15Z

Code

← Older revision		Revision as of 20:58, 18 March 2013
Line 149:		Line 149:
	= Appendix I - Code =		= Appendix I - Code =

	~~==Code==~~
	[[File:SchmittWerner.zip]]		[[File:SchmittWerner.zip]]

imported>Projects221: /* Appendix II - Work partition (if a group project) */

2013-03-18T20:58:03Z

Appendix II - Work partition (if a group project)

← Older revision		Revision as of 20:58, 18 March 2013
Line 152:		Line 152:
	[[File:SchmittWerner.zip]]		[[File:SchmittWerner.zip]]

	= Appendix II - Work partition ~~(if a group project)~~ =		= Appendix II - Work partition =

	Drew Schmitt: FFT, Metrics, SVM, Thresholding		Drew Schmitt: FFT, Metrics, SVM, Thresholding

	Kurt Werner: Raster Scanner, FFT, PCA, Metrics		Kurt Werner: Raster Scanner, FFT, PCA, Metrics

imported>Projects221: /* Future Work */

2013-03-18T20:57:23Z

Future Work

← Older revision		Revision as of 20:57, 18 March 2013
Line 133:		Line 133:
	= Future Work =		= Future Work =

	Future research should focus on abstracting the metric thresholding method ~~into~~ a higher dimensional space. ~~Additional~~ reasonably orthogonal high-performing metrics allowed SVM to perform better than our proposed approach. However, the lack of direct tuning of the false positive rate kept this method from being a useful technique. If the false positive rate could be controlled in an SVM setting, it is theorized that the classification results would improve considerably.		Future research should focus on abstracting the metric thresholding method developed in this paper to a higher dimensional space. It was experimentally determined that additional reasonably orthogonal high-performing metrics allowed SVM to perform better than our proposed approach. However, the lack of direct tuning of the false positive rate kept this method from being a useful technique. If the false positive rate could be controlled in an SVM setting, it is theorized that the classification results would improve considerably.

	= References - Resources and related work =		= References - Resources and related work =