Contrast Enhancement- Based Image Forgery Detection

Transcription

1 Contrast Enhancement- Based Image Forgery Detection Anju s Prasad 1, S. Antony Mutharasan 2 1 (CSE,Anna University,Tirunelveli,India,anjusp12@gmail.com) 2 ( CSE,Anna University,Tirunelveli,India,soby5882@gmail.com) Abstract With the rapid development of digital media editing techniques, digital image manipulation becomes rather convenient and easy. With the increasing applications of digital imaging, different types of software tools are introduced for image processing. They are used to combine two images to make it look real or objects can be added or deleted. The manipulation techniques include deletion of details, insertion of details, combining multiple images and false captioning. Detecting these image manipulations has become an important problem. To avoid these problems SVM classifier is proposed which have similar functional form to neural networks. Image, texture and pixel value based features are extracted and analyzed from the images. Then hash values are calculated for these features. The process consists of two phases which are training phase and a testing phase. SVM classifier is trained with a set of images and used to classify the images as genuine or forged. SVM supports both regression and classification tasks and can handle multiple continuous and categorical variables. This method reduces the time and computational complexity. Keywords SVM, digital forensics, image forgery, contrast enhancement, composite image 1.INTRODUCTION With the rapid development of digital media editing techniques, digital image manipulation becomes rather convenient and easy. While it benefits to legal image processing, malicious users might use such innocent manipulations to tamper digital photograph images. Currently, image forgeries are widespread on the Internet and other security related applications such as surveillance and recognition that utilize images are therefore impacted. The event and scene information delivered in images might become no longer believable. In the applications such as law enforcement and news recording, it is also necessary to verify the originality and authenticity of digital images, and make clear the image manipulation history to get more information. To circumvent such a problem, digital forensic techniques have been proposed to blindly verify the integrity and authenticity of digital images. A set of previous works deal with image manipulation detection by classifier-based approaches. Three categories of statistical features including binary similarity, image quality and wavelet statistics were developed. Swamina than proposed to estimate both in-camera and post-camera operation fingerprints for verifying the integrity of photographs. Cao designed a new ensemble manipulation detector to simultaneously detect a wide range of manipulation types on local image patches. Fan et al. proposed to correlate statistical noise features with exchangeable image file format header features for manipulation detection. Although such techniques could detect if manipulation occurred, they fail to determine which specific type of manipulation was enforced. There also exist another category of forensic techniques which focus on detecting specific image manipulations. Since each manipulation typically leaves behind unique fingerprints on images, it is feasible to design individual tests to identify each type of enforced manipulation. The manipulation-specific detection techniques can help recover the image processing history. The prior works focus on detecting different types of alterations, which can be broadly divided into two categories: 1) non-content-changing operations including resampling, compression, sharpening filtering, contrast enhancement and median filtering; 2) content-changing operations, i.e., splicing and composition.note that the prior contrast enhancement forensic algorithms work well under the assumption that the gray level histogram of an unaltered image exhibits a smooth contour. However, digital images are often stored in the JPEG format and even heavily compressed with a middle/low quality factor (Q) in real applications, such as the Internet and mobile phones. It is well-known that the low quality JPEG compression usually generates blocking artifacts, which might cause un smoothness and even locally dense peak bins in the gray level histogram. In such a scenario, the existing approaches fail to detect contrast enhancement in the previously low quality JPEG-compressed images, since the assumption of smoothness becomes dissatisfied. To solve such a problem, we propose to detect the global contrast enhancement not only in uncompressed or high quality JPEG-compressed images, but also in low quality ones. The main strategy relies on the blind identification of zero-height gap bins. Besides global contrast enhancement, the detection of local contrast enhancement is also significant. A valuable application is to identify the cut-and-paste type of forgery images, in which the contrast of one source object region is enhanced to match the rest. Although the composite image created by enhancing single source region could be identified, those enhanced in both source regions may not. In this paper, a new method is proposed to identify both single source enhanced and both source enhanced composite images. Peak/gap pattern of the pixel value mapping applied to each source region is self-learned from the detected block wise peak/gap positions. Then composition boundary is located by detecting the inconsistency between the position vectors in different regions. An SVM is one of many algorithms for supervised learning. Generally, an SVM is a linear classification algorithm that ensures us that the distance between the decision line (discriminator) and the closest example in the training set is maximized. Support Vector Machines are based on the concept of decision planes that define decision Volume: 02 Issue:

2 boundaries. A decision plane is one that separates between a set of objects having different class memberships. A schematic example is shown in the illustration below. In this example, the objects belong either to class GREEN or RED. The separating line defines a boundary on the right side of which all objects are GREEN and to the left of which all objects are RED. Any new object (white circle) falling to the right is labeled, i.e., classified, as GREEN Figure 1.1 Support Vector Machine The above is a classic example of a linear classifier, i.e., a classifier that separates a set of objects into their respective groups (GREEN and RED in this case) with a line. Most classification tasks, however, are not that simple, and often more complex structures are needed in order to make an optimal separation, i.e., correctly classify new objects (test cases) on the basis of the examples that are available (train cases). This situation is depicted in the illustration below. Compared to the previous schematic, it is clear that a full separation of the GREEN and RED objects would require a curve (which is more complex than a line). Classification tasks based on drawing separating lines to distinguish between objects of different class memberships are known as hyper plane classifiers. Support Vector Machines are particularly suited to handle such tasks. 2.PREVIOUS WORKS Digital Photo images are everywhere, on the covers of magazines, in newspapers, in courtrooms, and all over the Internet. Expose to throughout the day and most of the time. Which images can be easily manipulate, need to be aware that seeing does not always simply believing. Then propose methodologies to identify such unbelievable photo images and succeeded to identify forged region by given only the forged image. Formats are additive tag for every file system and contents are relatively expressed with extension based on most popular digital camera uses JPEG and other image formats. Design algorithm running behind with the concept of abnormal anomalies and identify the forgery regions. Image forgeries are widespread on the Internet and other security-related application such as surveillance and recognition that utilize images are therefore impacted. The event and scene information delivered in images might become no longer believable. In the applications such as law enforcement and news recording, it is also necessary to verify the originality and authenticity of digital images, and make clear the image manipulation history to get more information. To circumvent such a problem, digital forensic techniques have been proposed to blindly verify the integrity and authenticity of digital images. A set of previous works deal with mainly five categories: pixel-based techniques detect statistical anomalies introduced at the pixel level format-based techniques leverage the statistical correlations introduced by a specific lossy compression scheme camera-based techniques exploit artifacts introduced by the camera lens, sensor or onchip post-processing physically-based techniques explicitly model and detect anomalies in the three dimensional interaction between physical objects, light, and the camera Geometric based techniques make measurements of objects in the world and their positions relative to the camera. 2.1 Contrast Enhancement Detection Algorithm Based on the above analyses, propose the contrast enhancement detection algorithm as follows. 1) Get the image s normalized gray level histogram h (x). 2) Detect the bin at k as a zero-height gap bin if it satisfies: h(kk) = 0 min{h(kk 1), h(kk + 1)} > ττ 1 kk+ωω1 xx=kk ωω1 h(xx) > ττ 2ωω1+1 (2.1) Here, the first sub-equation assures that the current bin is null. To define a gap bin, the second sub-equation keeps two neighboring bins larger than the threshold τ, as shown. To exclude the zero-height gap bins which may be incorrectly detected in histogram trail-ends, the average of neighboring (2w1 + 1) bins should be larger than τ, as constrained by the third sub-equation. Experiments show that w1 = 3andτ = are appropriate. Note that focus on the detection of isolated zero-height gap bins but not connected bins, which are rarely present in the middle of histograms. 3) Count the number of detected zero-height gap bins, denoted by N g. If it is larger than the decision threshold, contrast enhancement is detected, else not. 3 SYSTEM ARCHITECTURE A novel algorithm is proposed to identify the sourceenhanced composite image created by enforcing contrast adjustment on either single or both source regions. The outline of technique is shown in Figure 3.1. Since positional distribution of the peak/gap bins incurred by contrast enhancement is unique to the involved pixel value mapping, such positional information could serve as fingerprinting feature for identifying different contrast enhancement manipulations. SVM classifier has similar functional form to neural networks. Image, texture and pixel value based features are extracted and analyzed from the images. Then hash values are calculated for these features. The process consists of two phases which are training phase and a testing phase. SVM classifier is trained with a set of images and used to classify the images as genuine or forged. SVM supports both regression and classification tasks Consistency Volume: 02 Issue:

3 Read input image between the peak/gap artifacts detected in different regions is checked for discovering composite images Block wise Peak/Gap Bins Location Normalized To locate composition, the test image is first divided into nonoverlapping blocks. For the i -th block, peak/gap bins in its gray level histogram are located as follows. Here and below, i = 1, 2,, Nb, where Nb is the number of divided blocks. Gap Gap Bins Location: The zero-height gap bins are detected.the position of detected gap bins is labelled as Vig= [Vig(0), Vig(1),, Vig(k),, V ig (255)] (3.1) Where Vig(k) = 1, if the bin at k is a gap; Vig(k) = 0, otherwise. Contrast Enhanced n Forgery image y Normal Peak Bins Location: Peak bins which behave as impulse noise can be located by median filtering. Specifically, the gap bins are first filled with the average of neighboring bins, and then median filtering is applied to the gap-filled histogram. The filtered histogram possesses a smooth contour. Lastly, peak positions are located by thresholding the difference between the gap-filled histogram and its filtered version. The histogram differences for the enhanced. Record the detected peak positions Vip= [Vip(0), Vip(1),, Vip(k),, Vip(255) (3.2 Where Vip(k) =1 refers to a peak Similarity Measure using reference vectors Block wise Gap & Peak Gap &Peak Similarity SVM Classifier To discriminate two source regions, first set a reference position vector for either one. Then each block can be classified by the similarity between its position vector and the reference one. It is reasonable to deem that the blocks with approximate similarity comes from the same source image. The reference position vector should not be selected from splicing boundary fortunately; the block with the largest number of zero-height gap bins is believed to locate within one source region. In boundary blocks, the interaction between the pixels from different source regions makes the number of zero height bins decrease. As a result, the reference gap position vector for its located source region. To measure the overall similarity between the gap position vectors, each gap involved pair should be investigated firstly. Since the histogram at the gray levels out of EDR does not deliver any effective peak/gap information left by contrast enhancement, the element pairs in the intersection of two EDRs are used to measure the similarity. As there exist three possible correspondences for a gap-involved pair. The ratio between the number of matched pairs and that of total gapinvolved pairs is defined as the similarity applied to the image region out of the reference block. Detect Tampering Region Performance analysis The narrow histogram without gap bins might carry with peak bins. As such the peak bins should also be exploited to identify mappings. The reference peak position vector is created by combining the peak position vectors which are more possible from the source region of gap vector. Such creation of position vector is reliable since the collected peak information is relatively accurate. Volume: 02 Issue:

4 3.3.3 Svm Classifier Support vector machines (SVMs) are a set of related supervised learning methods that analyze data and recognize patterns, used for classification and analysis. The standard SVM takes a set of input data and predicts, for each given input, which of two possible classes the input is a member of, which makes the SVM a non-probabilistic binary linear classifier. Since an SVM is a classifier, then given a set of training examples, each marked as belonging to one of two categories, an SVM training algorithm builds a model that predicts whether a new example falls into one category or the other. Intuitively, an SVM model is a representation of the examples as points in space, mapped so that the examples of the separate categories are divided by a clear gap that is as wide as possible. New examples are then mapped into that same space and predicted to belong to a category based on which side of the gap they fall on. A support vector machine constructs a hyper plane or set of hyper planes in a high or infinite dimensional space, which can be used for classification, regression, or other tasks. Intuitively, a good separation is achieved by the hyper plane that has the largest distance to the nearest training data points of any class (socalled functional margin), since in general the larger the margin the lower the generalization error of the classifier. Whereas the original problem may be stated in a finite dimensional space, it often happens that in that space the sets to be discriminated are not linearly separable. For this reason it was proposed that the original finite dimensional space be mapped into a much higher dimensional space, presumably making the separation easier in that space. SVM schemes use a mapping into a larger space so that cross products may be computed easily in terms of the variables in the original space, making the computational load reasonable. The cross products in the larger space are defined in terms of a kernel function K(x,y) selected to suit the problem. The hyper planes in the large space are defined as the set of points whose cross product with a vector in that space is constant. The vectors defining the hyper planes can be chosen to be linear combinations with parameters α i of images of feature vectors that occur in the data base. 4 MODULE DESCRIPITION 1. Contrast Enhancement Detection 2. Peak/Gap bins detection 3. Peak/Gap similarity Measure 4. SVM Composite Detection 5. Performance Analysis 1. Contrast enhancement detection 1) Get the image s normalized gray level histogram h (x). 2) Detect the bin at k as a zero-height gap bin if it satisfies: h(kk) = 0 min{h(kk 1), h(kk + 1)} > ττ (4.1) 1 kk+ωω1 h(xx) > ττ 2ωω1+1 xx=kk ωω1 Here, the first sub-equation assures that the current bin is null. To define a gap bin, the second sub-equation keeps two neighboring bins larger than the threshold τ. To exclude the zero-height gap bins which may be incorrectly detected in histogram trail-ends, the average of neighboring (2w1 + 1) bins should be larger than τ, as constrained by the third subequation. 3) Count the number of detected zero-height gap bins, denoted by N g. If it is larger than the decision threshold, contrast enhancement is detected, else not. 2. Peak/gap bins detection The test image is first divided into no overlapping blocks. For the i -th block, peak/gap bins in its gray level histogram are located. Here, i = 1, 2,, N b, where N b is the number of divided blocks. The position of detected gap bins is labelled as VV ii gg = [VV ii gg (0), VV ii gg (1)., VV ii gg (kk),., VV ii gg (255) if VV ii gg (kk) = 1 the bin at k is a gap, VV ii gg (kk) = 0 otherwise. The gap bins are first filled with the average of neighboring bins, then median filtering is applied to the gap-filled histogram. Lastly, peak positions are located by thresholding the difference between the gap-filled histogram and its filtered version. The detected peak positions as V i p=[ V i p (0), V i p (1),, V i p (k),. V i p (255)]. To further decrease detection errors, the extracted peak/ gap positions are corrected by retaining the co-existing peak/ gap positions in most blocks. Specifically, apply a threshold-based binarization to Cg NN = bb ii ii=1 VV gg /NN bb. The detected co-existing gap positions are recorded as V g =[V g ( 0 ), V g ( 1 ),, V g ( k ),, V g (255)], where V g ( k ) = 1, if C g ( k ) is larger than the threshold; V g ( k ) = 0, otherwise. To eliminate the gap bins which might not be caused by contrast enhancement, the corrected gap position vector, V i gc, is generated Hadamard product, Similarly, the corrected peak position vectorv i pc can also be obtained. 3. Peak/gap similarity measure To discriminate two source regions, first set a reference position vector for either one. Then each block can be classified by the similarity between its position vector and the reference one. To measure the overall similarity between the gap position vectors V i gc and V gr, each gap-involved pair V i gc ( k ) and V gr ( k ) should be investigated firstly. Here, the element pairs in the intersection of two EDRs are used to measure the similarity. There exist three possible correspondences for a gap-involved pair. The reference peak position vector V pr is created by combining the peak position vectors which are more possible from the source region of V pr, namely S 1.Such creation of V pr is reliable since the collected peak information is relatively accurate. The similarity between V i pc and V pr, marked as m i p, is defined in the same form by replacing the gap variables with the corresponding peak ones. If no peak-involved pair exists in EDR intersection, mark m i p = SVM Composite detection Extract each block gap similarity & peak similarity based on ground truth image, train the SVM network. Here design the two class svm classifier,composite versus normal. In training update the weight value of svm. g(x)=w^tx+w0 (4.2) Volume: 02 Issue:

5 where w^t,w0 are the weight values, x is the input. After training predict the output class table based on each block features. 5. Performance analyses In performance analysis, the effectiveness and efficacy of the proposed techniques is compared with the existing techniques. 5.CONCLUSIONS In this paper, propose SVM classifier which has similar functional form to neural networks. Image, texture and pixel value based features are extracted and analyzed from the images. Then hash values are calculated for these features. The process consists of two phases which are training phase and a testing phase. SVM classifier is trained with a set of images and used to classify the images as genuine or forged. SVM supports both regression and classification tasks and can handle multiple continuous and categorical variables. The composition boundary was accurately located by detecting the inconsistency between detected block wise peak/gap positional distributions. In the proposed method reduces the time and computational complexity. As future work to this project, apply an optimal thresholding technique to improve the Detection accuracy which is low due to fixed threshold and also highly noise sensitive for real time applications. REFERENCES 1. T. Arici, S. Dikbas, and Y. Altunbasak, A histogram modification framework and its application for image contrast enhancement, IEEE Trans. Image Process., vol. 18, no. 9, pp , Sep M. Barni, M. Fontani, and B. Tondi, A universal technique to hide traces of histogram-based image manipulations, in Proc. ACM Multimedia Security Workshop, Coventry, England, 2012, pp S. Battiato, A. Bosco, A. Castorina, and G. Messina, Automatic image enhancement by content dependent exposure correction, EURASIP J. Appl. Signal Process., vol. 12, pp , Jan S. Bayram, I. Avcubas, B. Sankur, and N. Memon, Image manipulation detection, J. Electron. Imag., vol. 15, no. 4, pp , T. Bianchi and A. Piva, Detection of non-aligned double JPEG compression based on integer periodicity maps, IEEE Trans. Inf. Forensics Security, vol. 7, no. 2, pp , Apr H. Cao and A. C. Kot, Manipulation detection on image patches using FusionBoost, IEEE Trans. Inf. Forensics Security, vol. 7, no. 3, pp , Jun G. Cao, Y. Zhao, R. Ni, and A. C. Kot, Unsharp masking sharpening detection via overshoot artifacts analysis, IEEE Signal Process. Lett., vol. 18, no. 10, pp , Oct G. Cao, Y. Zhao, and R. Ni, Forensic estimation of gamma correction in digital images, in Proc. 17th IEEE Int. Conf. Image Process., Hong Kong, 2010, pp C. Chen, J. Ni, and J. Huang, Blind detection of median filtering in digital images: A difference domain based approach, IEEE Trans. Image Process., vol. 22, no. 12, pp , Dec J. Fan, H. Cao, and A. C. Kot, Estimating EXIF parameters based on noise features for image manipulation detection, IEEE Trans. Inf. Forensics Security, vol. 8, no. 4, pp , Apr H. Farid, Image forgery detection, IEEE Signal Process. Mag., vol. 26, no. 2, pp , Mar P. Ferrara, T. Bianchiy, A. De Rosaz, and A. Piva, Reverse engineering of double compressed images in the presence of contrast enhancement, in Proc. IEEE Workshop Multimedia Signal Process., Pula, Croatia, Sep./Oct. 2013, pp S.-C. Huang, F.-C. Cheng, and Y.-S. Chiu, Efficient contrast enhancement using adaptive gamma correction with weighting distribution, IEEE Trans. Image Process., vol. 22, no. 3, pp , Mar M. Kirchner and J. Fridrich, On detection of median filtering in digital images, in Proc. SPIE, Electronic Imaging, Media Forensics and Security II, vol. 7541, San Jose, CA, USA, Jan. 2010, pp D. Mahajan, R. Ramamoorthi, and B. Curless, A theory of frequency domain invariants: Spherical harmonic identities for BRDF/lighting transfer and image consistency, IEEE Trans. Pattern Anal. Mach. Intell., vol. 30, no. 2, pp , Feb B. Mahdian and S. Saic, A bibliography on blind methods for identifying image forgery, Image Commun., vol. 25, no. 6, pp , Jul W. B. Pennebaker and J. L. Mitchell, JPEG Still Image Data Compression Standard. New York, NY, USA: Van Nostrand, C. Popescu and H. Farid, Exposing digital forgeries by detecting traces of resampling, IEEE Trans. Signal Process., vol. 53, no. 2, pp , Feb Volume: 02 Issue: