- Email: [email protected]
Forgery detection in digital images via discrete wavelet and discrete cosine transforms
ARTICLE IN PRESS
JID: CAEE
[m3Gsc;April 5, 2017;5:46]
Computers and Electrical Engineering 0 0 0 (2017) 1–11
Contents lists available at ScienceDirect
Computers and Electrical Engineering journal homepage: www.elsevier.com/locate/compeleceng
Forgery detection in digital images via discrete wavelet and discrete cosine transformsR Khizar Hayat a,b,∗, Tanzeela Qazi a a b
COMSATS Institute of Information Technology, University Road, Abbottabad 22060, Pakistan Computer Science Section (DMPS), College of Arts and Sciences, University of Nizwa, Sultanate of Oman
a r t i c l e
i n f o
Article history: Received 5 January 2016 Revised 13 March 2017 Accepted 14 March 2017 Available online xxx Keywords: Image tampering Image forgery Discrete Wavelet Transform(DWT) Discrete Cosine Transform (DCT) Image forgery detection
a b s t r a c t When an image is forged through some sophisticated software editing tool, with imperceptibility being the goal, the idea is not to leave any observable trace that may help to distinguish the forged image from the original one–at least with a naked eye. We believe that no matter which type of forgery is employed, there ought to be imperfections in the tampered image that may eventually prove its fakery. Even, forging an image with noble intentions, needs a reasonable amount of care. With this in perspective, we present a forgery detection method that depends on the discrete wavelet transform (DWT) as well as the discrete cosine transform (DCT) for feature reduction. The DCT is applied to the individual blocks obtained after dividing the DWTed image. The blocks are then compared on the basis of correlation coefficients. A mask-based tampering method is also developed as part of the experiments in order to test the detection method. The method shows interesting results when compared to two methods from the literature. © 2017 Elsevier Ltd. All rights reserved.
1. Introduction In today’s digital world, many different tools are used for the manipulation of images. In most cases, the altered image may not have any apparent clues regarding the original, especially if the latter is not easily accessible. Due to these sophisticated image editing software tools, the authenticity of a given image is always questionable and may give rise to many issues. No longer can the authenticity and integrity be guaranteed. The main problem is that it undermines the credibility of digital image as photographic evidence. Another important issue concerning digital fakery is the relative ease to use the tools and graphics algorithms, leading to serious vulnerabilities, that cast doubt on the integrity of digital images [1,2]. These problems bring into the fore the need to develop such tools and techniques that can easily discriminate the natural images from the tampered or synthetically generated images. Digital image forensics is a field that analyzes images of a particular scenario, in order to establish the credibility and authenticity (or otherwise), through a variety of means. It is fast becoming a popular field because of its potential applications in many domains like intelligence, sports, legal services, news reporting, medical imaging and insurance claim investigations [1,2]. With the ever increasing reliance on digital media, the need to ensure its authenticity and trustworthiness is of vital importance. The creation and manipulation of digital images, with no obvious tampering, is becoming easier and easier with
R ∗
Reviews processed and recommended for publication to the Editor-in-Chief by Guest Editor Dr. E. Cabal-Yepez. Corresponding author. E-mail address: [email protected] (K. Hayat).
http://dx.doi.org/10.1016/j.compeleceng.2017.03.013 0045-7906/© 2017 Elsevier Ltd. All rights reserved.
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE 2
ARTICLE IN PRESS
[m3Gsc;April 5, 2017;5:46]
K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
each passing day. Unfortunately, the counter efforts are undermined by the lacking of techniques to ascertain the origin or potential integrity of digital images. Research in digital image forensic needs to be aggressive and based on solid reasoning in order to uncover the actual facts while keeping pace with the development of image editing tools. In this paper we present a forgery detection method that is applicable in the case of copy/move forgery. The proposed transform domain technique relies on both the discrete wavelet transform (DWT) and the discrete cosine transform (DCT). The purpose is to reduce the features by first applying DWT to get the approximate sub-band. This is followed by dividing the latter to fixed sized square blocks and then applying DCT to each block separately, in order to reduce the features further. After lexicographically sorting the reduced set of features, correlation coefficients are used to judge the similarity of blocks. The rest of the paper is arranged as follows. Section 2 presents a concise survey of the related literature. This is followed by the presentation of the proposed method in Section 3. The simulation results are illustrated in Section 4 in somewhat detail. Section 5 concludes the paper. 2. Related work In early 1840’s, Hippolyte Bayrad created the first known fake image in which he was shown committing suicide1 . About two decades later, another fake image was created in which the head of Lincoln (US President) was shown on the body of John Calhoun, a Southern politician [3]. In the subsequent years, there had been a steady growth in the population of such tampered images. With the advent of color photography and subsequent coming of digital age, on the fly tampering was just a fait accompli. Today many advanced software and editing tools are available that can rather seamlessly forge images through a variety of techniques . Due to such forgeries, photography has lost its innocence. Image forgeries can be classified as [2]: • • •
Copy/move forgery, Image splicing and Image retouching.
The techniques from each of these categories can be implemented additionally via a) active or b, passive approaches [4]. The active approaches are mostly concerned with the data hiding techniques, such as digital watermarking/copyrighting, wherein prior information is considered essential and integral to the process. The passive approaches do not require any prior information about the original image . The passive blind techniques, where the analyzer has just the final product at his/her disposal, provide a solution to identify image alterations without relying on the insertion of an extrinsic data or digital signatures for image authentication. Forgery detection methods are broadly categorized as visual and statistical methods. Visual methods are based on the visual clues and sometimes require no hardware or software tools. It only detects, intelligently, the inconsistencies and light information of an image. In contrast, the statistical methods analyze pixel values of image and are hence more robust and convincing. The passive blind forgery detection scheme, outlined in [5], is based on content adaptive quantization table estimation. This technique is used for the detection of different types of forgeries, i.e. copy/move, splicing and synthetics. The accuracy rate is claimed to be high as compared to other techniques. It is claimed to be robust against the JPEG compression. For a detailed account of the methods on forgery detection, the readers can consult our earlier work on the subject [6]. Several surveys and feature analysis studies are available on copy/move forgery detection [7,8]. Many copy/move forgery detection techniques, proposed in the literature, exhibit good results but at the cost of relatively high computational time . Most of the copy/move forgery detection rely on block-wise comparison and the risk is that, “Similar but Genuine Objects (SGO)” may be treated as copied objects, as explained in [9]. Ideally, in a given image, all SGOs must be accounted for while subjecting it to copy/move forgery detection. The copy/move forgery detection technique of [10] is based on the DCT. The DCT coefficients of each block are selected to represent specific blocks. This is followed by lexicographic sorting of the four features of each block in order to check the similarity measure on the basis of a threshold value. The method is claimed to be robust against copy/move forgery. The downside is that it also detects wrong similar blocks and is sensitive to the addition noise or blurring. In [11], the authors detect copy/move forgery with the quantized DCT coefficients based block matching. The limitations of their technique include false identification of a few copied areas and low reliability with small copied images. The method in [12] divides the image into overlapping blocks and computes the DCT coefficients. By using the signs of the DCT coefficients, binary feature vectors are created. The latter are matched using the coefficient of correlation. A related method [13] employs principal component analysis (PCA) for feature reduction. In [14], the authors propose to treat the overlapping blocks to Local Binary Pattern (LBP) before applying the DCT and subsequent lexicographic sorting. Invariance to affine transformation, especially the scale, is an important aspect of copy/move forgery detection [15]. The method in [16] employs scale invariant feature transform (SIFT) in combination to DCT. The method outlined in [17] attempts to use statistical moments, computed from DCT quantized coefficients; for scale invariance, perhaps.
1
http://www.fourandsix.com/photo- tampering- history/?currentPage=8
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE
ARTICLE IN PRESS K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
[m3Gsc;April 5, 2017;5:46] 3
Discrete Wavelet Transform (DWT) is a popular transform thanks to its localized nature and its ability to compact most of the image information into the lowest energy sub-band that is dyadically reduced in size in proportion to the image size. Hence, rather than the suspect image, its lowest energy sub-band can be subjected to forensic analysis in order to reduce the complexity – e.g. a level-2 sub-band would have sixteen times less coefficients to analyze. In addition, DWT may enable the extraction of very good and robust features for comparisons. A DWT based method [18], first exhaustively searches for the identification of matching blocks and then uses phase correlation for the detection of the copied regions. However, the technique gives poor results if the copied region is slightly scaled or rotated [2,18]. In [19], pixel matching and DWT techniques are utilized to reduce the dimensions. Moreover, phase correlation is used for the detection steps in the copied and pasted regions. To improve the forgery localization, mathematical morphology is employed for the connected regions. The above mentioned technique has low complexity and exhibits robustness against the post processing of the copied regions. However, the performance depends on the scene of the copy/move image. Another copy/move forgery detection algorithm, for color images, is on the basis of sensor pattern noise (SPN) [20]. Pattern noise is extracted by using the wavelet based Wienerâ;;s denoising filter. The features are selected on the basis of signal to noise ratio, information entropy, variance of pattern noise and average of energy gradient of the extracted image. The method is shown to be robust against geometric transformations (rotation and scaling), noise and JPEG compression. The blind forensic technique for copy/move forgery detection, reported in [21] , uses DWT to reduce dimensions and to generate overlapping blocks from the compressed image. Thereafter, the blocks are sorted lexicographically and phase correlation is used for checking block-wise similarities. The technique reduces the duration of the detection process. Another DWT method, by the same authors and proposed in [22], demonstrate a higher detection rate compared to the technique detailed in [11]. The technique consists of two phases and uses multiresolution characteristics of the wavelets transform to reduce the dimensions. Thereafter, overlapping blocks of fixed sizes are lexicographically sorted and checked through the similarity measures. The technique exhibits robustness even if the forged region is retouched further. In [23], a Dyadic wavelet transform (DyWT) based method is proposed for blindly detecting copy/move forgery. The suspect image is decomposed using the DyWT and the approximate (LL1) and detailed (HH1) sub-bands are extracted. The decision is made on the basis of similarity measures between the blocks of sub bands LL1 and HH1. Zimba and Xingming [24] combine DWT and PCA to detect the copy/move forgery but the method shows poor robustness. Many works, in splicing digital images, are based on the transform domain. The approach elaborated in [25] is developed to circumvent JPEG resizing, and relies on the correlation of the surroundings values of the DCT coefficients with the help of a support vector machine (SVM). The authors argue that the scale factors may seriously affect the performance of the methods. Moreover, the detection is far from satisfactory in highly complex situations. The method in [26] employs Markov features and DCT for feature extraction, which are subsequently reduced by PCA, followed by the use of a SVM for classification. Sun et al. present a wavelet domain method on the basis of the natural image statistical model. Generalized Gaussian model (GGD) is employed for estimating the parameters and features from each wavelet sub-band. Due to the high detection accuracy and simplicity, the above mentioned technique can be utilized for many applications [27]. The method of Jing et al. [28] blindly looks for side effects resulted in the image after being tampered. The authors refer to these side effects as the eclosion traces for their resemblance with the metamorphosis in insects. Dual tree wavelet transform is used for the image decomposition in order to remove the noise attributed to the change in the texture. Based on the gathered information, the image can be reconstructed. This technique provides better accuracy when compared to the technique described in [29] that focuses on the in-harmonic points, which represent the pixels having different intensities from their neighbors. In the context of noise inconsistency, Mahdian and Saic have used the high resolution wavelet coefficients for the estimation of noise [30]. Obviously, the detection rate of such a technique is not satisfactory when the noise ratio is low. Moreover, if the variances of isolated regions in an image are equal, then the method fails to detect the forgery. In essence, the existing methods of splicing detection have high accuracy rate but most of the methods have time complexity issues and validity for geometric transformations is not that satisfactory. 3. Proposed work In order to detect image forgery, we propose a detection method that is based on the DWT and DCT. First, the lowest frequency sub-band or approximate coefficients are extracted by applying a DWT to the image. This is followed by the block-wise application of DCT to get the energy rich few coefficients. It will not be out of place here to briefly describe the motivation to use both DWT and DCT. Our approach relies on the division of the image to overlapping blocks for comparison, with a sliding factor of one pixel. Hence, in the absence of any padding, for a M × N image, there will be a total of (M − m + 1 ) × (N − m + 1 ) blocks of size m × m. This may lead to a huge number of comparisons, as each block must be compared with every other block. The situation is exacerbated if the image is large. As most of the information in an image is redundant, we believe the better way is to reduce the size of the image without losing most of the important information. The best way do so is to pass the image to frequency domain which would concentrate most of the energy in a few low frequency coefficients. At the same time, we do not want to lose the localization information, as the image has to be subsequently divided to overlapping blocks. With this in view, DWT was chosen as the transform to reduce the image size without compromising the localization information. DWT performs far better than DCT in getting an approximation of the image, especially in the absence of blocks. In the subsequent step, DCT is used because we are dealing with small blocks and for small blocks DCT is a better choice to compact the information. Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
ARTICLE IN PRESS
JID: CAEE 4
[m3Gsc;April 5, 2017;5:46]
K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
Fig. 1. Block diagram representation of proposed Forgery detection method.
Step by step process of the proposed detection method (Fig. 1) is as follows. 1. Consider gray scale image as input, if the image in RGB format convert to YCbCr and use Y as gray-scale image. The image is then subjected to necessary pre-processing, like noise removal/Gaussian smoothing and normalization. 2. The preprocessed image is then subjected to level l DWT to get 3l + 1 sub-bands. The value of l is proportional to the dimensions of the image. The larger the image, larger should be l and vice versa. However, the value of l seldom exceeds 5 and that too for very large images. The lowest frequency (high energy) sub-band LLl is chosen for further processing. The reduction in feature set size is dyadic and for a square image of size n × n, the lowest frequency sub-band has a size in the order of nl × nl . In other words, a level-3 DWT will give you approximate coefficients amounting to 1.56% of 2 2 the total. 3. The approximate lowest energy sub-band from the last step is divided into fixed size overlapping blocks. Although the DWT may have considerably reduced the number of coefficients, the number may still be large in face of the fact that we are dealing with overlapping blocks. If we continue the dyadic example from the preceding step, a sub-band of size n × nl may result in ( nl − m + 1 ) × ( nl − m + 1 ) square blocks of m × m coefficients, where m is the block size. Hence l 2
4.
5.
6.
7.
2
2
2
total number of coefficients, to deal with, are still ( nl − m + 1 )2 × m2 , which is a large set. 2 DCT is applied to each individual block in order to get the most energetic of coefficients from each block. Extract the most important (read energetic) features from each block in the form of a row vector, by following a zigzag order while starting from the DC coefficient. Note that we have now as many row vectors as were the number of blocks. Sort the block features, lexicographically, so that similar features become closer to each other. Note that lexicographic sorting is similar to that of dictionary, except with the additional criteria of size; sort by size first and then among the same size sort like a dictionary. (Fig. 2) Find the correlation coefficients between all the row vectors. Note that for our example from Step 3, we had k = ( nl − 2 m + 1 ) × ( nl − m + 1 ) blocks converted to as many vectors in step 4. Since each block i need to be compared with 2 each block j (j = i) to get the correlation coefficients of the form ρ ij , we end up with a total of k(k − 1 )/2 correlation coefficients. Note that −1 ≤ ρi j ≤ 1. Finally map the duplicated blocks by comparing all the ρ ij to a positive threshold ρ 0 . If for two blocks indexed i and j, |ρ ij | < ρ 0 , then mark them as duplicated. The marked blocks, if any, are indicators of the copy/move forgery.
It can be seen that the purpose of DWT is to reduce the image features to manageable size and then divide the latter to block that are made manageable by DCT application. Even then the number of coefficients may become very large if block sizes are not chosen with care. Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE
ARTICLE IN PRESS K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
[m3Gsc;April 5, 2017;5:46] 5
Fig. 2. Block diagram representation of image forgery by mask method.
4. Experimental results The proposed technique has been implemented and tested with images of various sizes and duplication at different regions therein. We have designed our experiments according to the following points: •
• •
Forge the image intelligently either using a state of the art image editing tool or by using the mask based method developed specially for this work. Apply the proposed detection method to the forged images. Compare the ensued results with those obtained by the application of forgery detection methods outlined in [10,24].
4.1. Image forgery by using the mask based method Alongside the photo-editing tool, we have developed our own editing method for a better idea on forgery detection through our method in comparison to others. The method consists of developing a binary multiplication mask in order to extract the part to be substituted as forgery in the original image. This method is applicable to copy/move forgery. The following steps are involved to forge the image: • • • •
Load the original RGB image, I1 . Select the region to be moved or copied. By using some image editing tool extract the specified image region that is to be placed. Convert this extracted region into a gray scale image.
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE 6
ARTICLE IN PRESS
[m3Gsc;April 5, 2017;5:46]
K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
Fig. 3. Image representation of image forgery by mask method.
•
•
Employ an edge detection technique, like Canny edge detection2 algorithm, in order to detect edges in the extracted region. Morphological processing is performed in order to complete the following tasks: 1. Dilate image to link the broken edges. 2. Hole/region filling
•
• •
The result would be a binary mask. Extract the object of interest by multiplying with the original, I1 . The result is an image (I2 ) having the object of interest with the rest being the background with zero intensity values. One can stretch the background to the size of the image to be forged (call it I3 - it is possible to have I1 = I3 ). In the correspondence between I3 and I2 , substitute a pixel of I3 with I2 , if the intensity of the latter is non-zero. Return the resultant forged I3 image with, apparently, no evidence of tampering.
Fig. 33 demonstrate the use of the mask based method for region duplication. We have tampered several images by using the above method and the image tampering results appeared seamless. The method is also capable to successfully forge via image splicing. Fig. 4 illustrates two more examples of forgeries4 . 4.2. Methods from literature used for comparison With regard to similarity to our method, we had chosen two methods for comparison to ours, in order to judge the effectiveness of the latter. The first method, by Cao et al. [10], will be hereafter called Cao’s method. The second method is by Zimba et al. [24] and will be hereinafter called Zimba’s method. 4.2.1. Cao’s method [10] This method is claimed to be robust against the detection of wrong similar blocks, having lower of feature vector dimensions and reduced time complexity. Step by step process of this detection method is as follows: 1. 2. 3. 4. 5.
The suspect RGB image is divided into fixed sized block. DCT is applied, followed by quantization, to get quantized coefficients. The blocks are represented by the circle block to extract fewer features in order to reduce the dimensions. Check the similar block pairs by correlation coefficients. Find the correct blocks and make a color map of duplicated block
4.2.2. Zimba’s method [24] This method detects image forgery by using a DWT to get the low frequency sub-band of the investigated image and then PCA is used to reduce the feature dimensions. Specifically, this method is used for the detection of copy/move image forgery or cloning. This algorithm consists of the following steps. 2 3 4
http://www.cse.iitd.ernet.in/∼pkalra/csl783/canny.pdf. Example image taken from www.hdwpics.com. Example 2 is after www.dreamstime.com and Example 3 is after www.hdwallpapers360.com.
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
ARTICLE IN PRESS
JID: CAEE
[m3Gsc;April 5, 2017;5:46]
K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
7
Fig. 4. Example forgeries. Table 1 Specification of detection algorithms.
1. 2. 3. 4. 5. 6. 7. 8.
Detection algorithm
Extraction domain
Block size
Feature dimension
Proposed method Cao’s method Zimba’s method
DWT-DCT Block based DCT DWT-PCA
8× 8 8× 8 8× 8
10 4 7
Consider a gray scale image, if the image is in RGB format then apply transform on each component separately. Apply the Wavelet transform to get the image details. Low frequency band is selected for further processing. Divide the image into overlapping blocks. For dimension reduction PCA â;; Eigen-value Decomposition is applied to get the row vectors. Row vectors are sorted lexicographically; as a result the similar vectors closest to each other. Compute the normalized shift vector between the adjacent block positions. Map the result on the basis of a preset threshold value.
Applicability of any copy/move forgery detection algorithm depends on the requirement that it should detect the duplicated regions which are either natural or forged. The specification of detection algorithms are shown in Table 1. 4.3. Experimental results of forgery detection The effectiveness of the forgery detection methods are usually gauged by the two measures given illustrated in the following equations:
r =| R
D|/|R|
w =| F − D | / | R |,
(1) (2)
where r is detection rate and w is false detection rate. R represents the actual copy/move tampered area and D is the detected area. Fig. 5 illustrates the detection rate accuracy of the proposed method in comparison to the reference methods for a number of tampered images from the testing set. The proposed method is superior in the sense that its average accuracy Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE 8
ARTICLE IN PRESS
[m3Gsc;April 5, 2017;5:46]
K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
Fig. 5. Detection rate accuracy of forged test images with different methods.
Table 2 Comparison table. Method
F.D.R
S.D
Detection accuracy
S.D
Proposed work Cao’s method Zimba’s method
6.47% 16.27% 10.31%
5.09 8.69 8.11
73.62% 72.77% 69.48%
10.74 9.85 10.23
is 73.62%, as against 72.77% and 69.48% for Cao’s and Zimba’s methods, respectively. Whilst the maximum accuracy for the proposed method (94.74%) is at par with Zimba’s method, it is far less for Cao’s method (87.88%). A better measure, to explain the spread, is standard deviation which was observed to be 10.74 as against 11.26 for Zimba’s method. The other method has a slightly better value of 9.85 but it suffers from lower average and maximum. Detection results of the proposed copy/move forgery algorithm are listed in Table 2, along with the results of the reference methods. The superiority of the proposed method is quite glaring with respect to the false detection rates (F.D.Rs), as can be observed from the graph in Fig. 6. With it’s high F.D.R., the Cao’s method is a poor performer; a fact also evident from its very high average F.D.R. – 16.27% as against 6.47% for the proposed. Zimba’s method is comparable to ours but still it’s average F.D.R is almost twice (10.31%). In fact the very low F.D.R. is the strength of our method, if read in combination with the standard deviation or S.D(5.09). For the two reference methods, the value is well above 8.
5. Conclusion The melange of DCT and DWT helped us achieving good results par rapport the two reference methods. The false detection rate was observed to be far less in comparison to the methods relying on either of the DCT and DWT. The low standard deviation of the false detection rate is like an icing on the cake. As of detection accuracy, the proposed method outperformed others by having highest average accuracy. The special multiplication mask based method, developed as part of experimentation, may serve in itself a seamless editing method. As can be observed from the example illustrations, the tampering is hard to be blindly noticed, i.e. only the presence of original image would establish the tampering. One additional advantage of the detection method was its viability for both copy/move and splicing based forgeries. All the aforementioned notwithstanding, there is a chance that the method may somehow under perform in the presence of occlusion. Same may be true of images with repeated patterns. In future, we intend to deal with these situations. We also plan to come up with a method to cover most of the contemporary forgeries. Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE
ARTICLE IN PRESS K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
[m3Gsc;April 5, 2017;5:46] 9
Fig. 6. False detection rate of forged test images with different methods.
References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26]
Mahdian B, Saic S. Blind methods for detecting image fakery. Aerosp Electr Syst Mag IEEE 2010;25(4):18–24. Shivakumar BL, Baboo SS. Detecting copy-move forgery in digital images: a survey and analysis of current methods. Glob J Comput Sci Technol 2010;10. Farid H. Digital doctoring: how to tell the real from the fake. Significance 2006;3(4):162–6. Zhang Z, Ren Y, Ping XJ, He ZY, Zhang SZ. A survey on passive-blind image forgery by doctor method detection. In: Proc. Seventh International Conference on Machine Learning and Cybernetics (ICMLC), Kunming, China; 2008. p. 3463–7. Lin G, Chang M, Chen Y. A passive-blind forgery detection scheme based on content-Adaptive quantization table estimation. Circ Syst Video Technol IEEE Trans 2011;21(4):421–34. Qazi T, Hayat K, Khan S, Madani S, Khan I, Kolodziej J, et al. Survey on blind image forgery detection. IET Image Proc 2013;7(7):660–70. Singhal N, Gandhani S. Analysis of copy-move forgery image forensics: a review. Int J Sig Process Image Process Pattern Recognit 2015;8(7):265–72. Qureshi MA, Deriche M. A bibliography of pixel-based blind image forgery detection techniques. Sig Process Image Commun 2015;39, Part A:46–74. Zhu Y, Ng TT, Shen X, Wen B. Revisiting copy-move forgery detection by considering realistic image with similar but genuine objects. CoRR 2016;abs/1601.07262. Cao Y, Gao T, Fan L, Yang Q. A robust detection algorithm for copy-Move forgery in digital images. Forensic Sci Int 2011. Fridrich J, Soukal D, Lukás˘ J. Detection of copy-move forgery in digital images. In: Proc. Digital Forensic Research Workshop; 2003. Singhal N, Gandhani S. Copy move forgery detection in contrast variant environment using binary DCT vectors. Int J Image Graph Sig Process 2015;6:38–44. Mahmood T, Nawaz T, Irtaza A, Ashraf R, Shah M, Mahmood MT. Copy move forgery detectiontechnique for forensic analysis in digital images. Math Probl Eng 2016;2016(8713202):13. Boz A, Bilge HS. Copy-move image forgery detection based on LBP and DCT. In: Proc. 24th Signal Processing and Communication Application Conference (SIU); 2016. p. 561–4. Warbhe AD, Dharaskar RV, Thakare VM. A scaling robust copy-paste tampering detection for digital image forensics. Procedia Comput Sci 2016;79:458–65. Kaur A, Shirma R. Copy-move forgery detection using DCT and SIFT. Int J Comput Appl 2013;70(7):30–4. Kaushik R, Bajaj RK, Mathew J. On image forgery detection using two dimensional discrete cosine transform and statistical moments. Procedia Comput Sci 2015;70:130–6. Myna AN, Venkateshmurthy MG, Patil CG. Detection of region duplication forgery in digital images using wavelets and log-polar mapping. In: Proc. International Conference on Computational Intelligence and Multimedia Applications (ICCIMA) - Volume 03; 2007. Zhang J, Feng Z, Su Y. A new approach for detecting copy-move forgery in digital images. In: Proc. 11th IEEE Singapore International Conference on Communication Systems (ICCS); 2008. p. 362–6. Peng F, Nie Y, Long M. A complete passive blind image copy-Move forensics scheme based on compound statistics features. Forensic Sci Int 2011. Khan S, Kulkarni A. An efficient method for detection of copy- Move forgery using discrete wavelet transform. Int J Comput Sci Eng (IJCSE) , 2010;05(02). Khan S, Kulkarni A. Detection of copy-move forgery using multiresolution characteristic of discrete wavelet transform. In: Proc. International Conference & Workshop on Emerging Trends in Technology (ICWET’11). ICWET ’11; 2011. p. 127–31. ISBN 978-1-4503-0449-8 Muhammad G, Bebis G. Blind copy move image forgery detection using dyadic undecimated wavelet transform. In: Proc. International Conference on Signal and Image Processing (ICSIP); 2010. Zimba M, Xingming S. DWT-PCA (EVD) based copy-move image forgery detection. Int J Digit Content Technol Appl (JDCTA) 2011;5(1):251–8. Liu Q., Sung H.A.. A new approach for JPEG resize and image splicing detection. In: Proc. First ACM workshop on Multimedia in forensics (MiFor’09). El-Alfy E-SM, Qureshi MA. Combining spatial and DCT based markov features for enhanced blind detection of image splicing. Pattern Anal Appl 2015;18(3):713–23.
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE 10
ARTICLE IN PRESS
[m3Gsc;April 5, 2017;5:46]
K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
[27] Sun S, Wu Q, Li G. Detection of image compositing based on a statistical model for natural images. Acta Autom Sin 2010;35:1564–7. [28] Jing T, Peng Y, Zhang F, Huo Y. Blind detection of digital forgeries using detection trace of eclosion. In: Proc. International Conference on Computational Intelligence and Software Engineering (CiSE’09); 2009. [29] Ying C, Yuping W. Exposing digital forgeries by detecting traces of smoothing. In: Proc. 9th International Conference for Young Computer Scientists (ICYCS); 2008. p. 1440–5. [30] Mahdian B, Saic S. Using noise inconsistencies for blind image forensics. Image Vis Comput 2009;27(10):1497–503. Special Section: Computer Vision Methods for Ambient Intelligence
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE
ARTICLE IN PRESS K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
[m3Gsc;April 5, 2017;5:46] 11
Khizar Hayat is heading the Computer Science Section of College of Arts and Sciences, University of Nizwa, Oman. He has also led the Computer Science Department of COMSATS Institute of Information Technology, Abbottabad, Pakistan. He received his PhD degree in 2009 from the University of Montpellier, France, while working at LIRMM. His preference areas are image processing and information hiding. Tanzeela Qazi received her MS degree in Computer Science from the COMSATS Institute of Information Technology (CIIT), Pakistan in 2011. She mainly did her thesis work at the Abbottabad Campus of CIIT. Her area of interest is digital image processing with special reference to image forgery detection.
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE
[m3Gsc;April 5, 2017;5:46]
Computers and Electrical Engineering 0 0 0 (2017) 1–11
Contents lists available at ScienceDirect
Computers and Electrical Engineering journal homepage: www.elsevier.com/locate/compeleceng
Forgery detection in digital images via discrete wavelet and discrete cosine transformsR Khizar Hayat a,b,∗, Tanzeela Qazi a a b
COMSATS Institute of Information Technology, University Road, Abbottabad 22060, Pakistan Computer Science Section (DMPS), College of Arts and Sciences, University of Nizwa, Sultanate of Oman
a r t i c l e
i n f o
Article history: Received 5 January 2016 Revised 13 March 2017 Accepted 14 March 2017 Available online xxx Keywords: Image tampering Image forgery Discrete Wavelet Transform(DWT) Discrete Cosine Transform (DCT) Image forgery detection
a b s t r a c t When an image is forged through some sophisticated software editing tool, with imperceptibility being the goal, the idea is not to leave any observable trace that may help to distinguish the forged image from the original one–at least with a naked eye. We believe that no matter which type of forgery is employed, there ought to be imperfections in the tampered image that may eventually prove its fakery. Even, forging an image with noble intentions, needs a reasonable amount of care. With this in perspective, we present a forgery detection method that depends on the discrete wavelet transform (DWT) as well as the discrete cosine transform (DCT) for feature reduction. The DCT is applied to the individual blocks obtained after dividing the DWTed image. The blocks are then compared on the basis of correlation coefficients. A mask-based tampering method is also developed as part of the experiments in order to test the detection method. The method shows interesting results when compared to two methods from the literature. © 2017 Elsevier Ltd. All rights reserved.
1. Introduction In today’s digital world, many different tools are used for the manipulation of images. In most cases, the altered image may not have any apparent clues regarding the original, especially if the latter is not easily accessible. Due to these sophisticated image editing software tools, the authenticity of a given image is always questionable and may give rise to many issues. No longer can the authenticity and integrity be guaranteed. The main problem is that it undermines the credibility of digital image as photographic evidence. Another important issue concerning digital fakery is the relative ease to use the tools and graphics algorithms, leading to serious vulnerabilities, that cast doubt on the integrity of digital images [1,2]. These problems bring into the fore the need to develop such tools and techniques that can easily discriminate the natural images from the tampered or synthetically generated images. Digital image forensics is a field that analyzes images of a particular scenario, in order to establish the credibility and authenticity (or otherwise), through a variety of means. It is fast becoming a popular field because of its potential applications in many domains like intelligence, sports, legal services, news reporting, medical imaging and insurance claim investigations [1,2]. With the ever increasing reliance on digital media, the need to ensure its authenticity and trustworthiness is of vital importance. The creation and manipulation of digital images, with no obvious tampering, is becoming easier and easier with
R ∗
Reviews processed and recommended for publication to the Editor-in-Chief by Guest Editor Dr. E. Cabal-Yepez. Corresponding author. E-mail address: [email protected] (K. Hayat).
http://dx.doi.org/10.1016/j.compeleceng.2017.03.013 0045-7906/© 2017 Elsevier Ltd. All rights reserved.
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE 2
ARTICLE IN PRESS
[m3Gsc;April 5, 2017;5:46]
K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
each passing day. Unfortunately, the counter efforts are undermined by the lacking of techniques to ascertain the origin or potential integrity of digital images. Research in digital image forensic needs to be aggressive and based on solid reasoning in order to uncover the actual facts while keeping pace with the development of image editing tools. In this paper we present a forgery detection method that is applicable in the case of copy/move forgery. The proposed transform domain technique relies on both the discrete wavelet transform (DWT) and the discrete cosine transform (DCT). The purpose is to reduce the features by first applying DWT to get the approximate sub-band. This is followed by dividing the latter to fixed sized square blocks and then applying DCT to each block separately, in order to reduce the features further. After lexicographically sorting the reduced set of features, correlation coefficients are used to judge the similarity of blocks. The rest of the paper is arranged as follows. Section 2 presents a concise survey of the related literature. This is followed by the presentation of the proposed method in Section 3. The simulation results are illustrated in Section 4 in somewhat detail. Section 5 concludes the paper. 2. Related work In early 1840’s, Hippolyte Bayrad created the first known fake image in which he was shown committing suicide1 . About two decades later, another fake image was created in which the head of Lincoln (US President) was shown on the body of John Calhoun, a Southern politician [3]. In the subsequent years, there had been a steady growth in the population of such tampered images. With the advent of color photography and subsequent coming of digital age, on the fly tampering was just a fait accompli. Today many advanced software and editing tools are available that can rather seamlessly forge images through a variety of techniques . Due to such forgeries, photography has lost its innocence. Image forgeries can be classified as [2]: • • •
Copy/move forgery, Image splicing and Image retouching.
The techniques from each of these categories can be implemented additionally via a) active or b, passive approaches [4]. The active approaches are mostly concerned with the data hiding techniques, such as digital watermarking/copyrighting, wherein prior information is considered essential and integral to the process. The passive approaches do not require any prior information about the original image . The passive blind techniques, where the analyzer has just the final product at his/her disposal, provide a solution to identify image alterations without relying on the insertion of an extrinsic data or digital signatures for image authentication. Forgery detection methods are broadly categorized as visual and statistical methods. Visual methods are based on the visual clues and sometimes require no hardware or software tools. It only detects, intelligently, the inconsistencies and light information of an image. In contrast, the statistical methods analyze pixel values of image and are hence more robust and convincing. The passive blind forgery detection scheme, outlined in [5], is based on content adaptive quantization table estimation. This technique is used for the detection of different types of forgeries, i.e. copy/move, splicing and synthetics. The accuracy rate is claimed to be high as compared to other techniques. It is claimed to be robust against the JPEG compression. For a detailed account of the methods on forgery detection, the readers can consult our earlier work on the subject [6]. Several surveys and feature analysis studies are available on copy/move forgery detection [7,8]. Many copy/move forgery detection techniques, proposed in the literature, exhibit good results but at the cost of relatively high computational time . Most of the copy/move forgery detection rely on block-wise comparison and the risk is that, “Similar but Genuine Objects (SGO)” may be treated as copied objects, as explained in [9]. Ideally, in a given image, all SGOs must be accounted for while subjecting it to copy/move forgery detection. The copy/move forgery detection technique of [10] is based on the DCT. The DCT coefficients of each block are selected to represent specific blocks. This is followed by lexicographic sorting of the four features of each block in order to check the similarity measure on the basis of a threshold value. The method is claimed to be robust against copy/move forgery. The downside is that it also detects wrong similar blocks and is sensitive to the addition noise or blurring. In [11], the authors detect copy/move forgery with the quantized DCT coefficients based block matching. The limitations of their technique include false identification of a few copied areas and low reliability with small copied images. The method in [12] divides the image into overlapping blocks and computes the DCT coefficients. By using the signs of the DCT coefficients, binary feature vectors are created. The latter are matched using the coefficient of correlation. A related method [13] employs principal component analysis (PCA) for feature reduction. In [14], the authors propose to treat the overlapping blocks to Local Binary Pattern (LBP) before applying the DCT and subsequent lexicographic sorting. Invariance to affine transformation, especially the scale, is an important aspect of copy/move forgery detection [15]. The method in [16] employs scale invariant feature transform (SIFT) in combination to DCT. The method outlined in [17] attempts to use statistical moments, computed from DCT quantized coefficients; for scale invariance, perhaps.
1
http://www.fourandsix.com/photo- tampering- history/?currentPage=8
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE
ARTICLE IN PRESS K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
[m3Gsc;April 5, 2017;5:46] 3
Discrete Wavelet Transform (DWT) is a popular transform thanks to its localized nature and its ability to compact most of the image information into the lowest energy sub-band that is dyadically reduced in size in proportion to the image size. Hence, rather than the suspect image, its lowest energy sub-band can be subjected to forensic analysis in order to reduce the complexity – e.g. a level-2 sub-band would have sixteen times less coefficients to analyze. In addition, DWT may enable the extraction of very good and robust features for comparisons. A DWT based method [18], first exhaustively searches for the identification of matching blocks and then uses phase correlation for the detection of the copied regions. However, the technique gives poor results if the copied region is slightly scaled or rotated [2,18]. In [19], pixel matching and DWT techniques are utilized to reduce the dimensions. Moreover, phase correlation is used for the detection steps in the copied and pasted regions. To improve the forgery localization, mathematical morphology is employed for the connected regions. The above mentioned technique has low complexity and exhibits robustness against the post processing of the copied regions. However, the performance depends on the scene of the copy/move image. Another copy/move forgery detection algorithm, for color images, is on the basis of sensor pattern noise (SPN) [20]. Pattern noise is extracted by using the wavelet based Wienerâ;;s denoising filter. The features are selected on the basis of signal to noise ratio, information entropy, variance of pattern noise and average of energy gradient of the extracted image. The method is shown to be robust against geometric transformations (rotation and scaling), noise and JPEG compression. The blind forensic technique for copy/move forgery detection, reported in [21] , uses DWT to reduce dimensions and to generate overlapping blocks from the compressed image. Thereafter, the blocks are sorted lexicographically and phase correlation is used for checking block-wise similarities. The technique reduces the duration of the detection process. Another DWT method, by the same authors and proposed in [22], demonstrate a higher detection rate compared to the technique detailed in [11]. The technique consists of two phases and uses multiresolution characteristics of the wavelets transform to reduce the dimensions. Thereafter, overlapping blocks of fixed sizes are lexicographically sorted and checked through the similarity measures. The technique exhibits robustness even if the forged region is retouched further. In [23], a Dyadic wavelet transform (DyWT) based method is proposed for blindly detecting copy/move forgery. The suspect image is decomposed using the DyWT and the approximate (LL1) and detailed (HH1) sub-bands are extracted. The decision is made on the basis of similarity measures between the blocks of sub bands LL1 and HH1. Zimba and Xingming [24] combine DWT and PCA to detect the copy/move forgery but the method shows poor robustness. Many works, in splicing digital images, are based on the transform domain. The approach elaborated in [25] is developed to circumvent JPEG resizing, and relies on the correlation of the surroundings values of the DCT coefficients with the help of a support vector machine (SVM). The authors argue that the scale factors may seriously affect the performance of the methods. Moreover, the detection is far from satisfactory in highly complex situations. The method in [26] employs Markov features and DCT for feature extraction, which are subsequently reduced by PCA, followed by the use of a SVM for classification. Sun et al. present a wavelet domain method on the basis of the natural image statistical model. Generalized Gaussian model (GGD) is employed for estimating the parameters and features from each wavelet sub-band. Due to the high detection accuracy and simplicity, the above mentioned technique can be utilized for many applications [27]. The method of Jing et al. [28] blindly looks for side effects resulted in the image after being tampered. The authors refer to these side effects as the eclosion traces for their resemblance with the metamorphosis in insects. Dual tree wavelet transform is used for the image decomposition in order to remove the noise attributed to the change in the texture. Based on the gathered information, the image can be reconstructed. This technique provides better accuracy when compared to the technique described in [29] that focuses on the in-harmonic points, which represent the pixels having different intensities from their neighbors. In the context of noise inconsistency, Mahdian and Saic have used the high resolution wavelet coefficients for the estimation of noise [30]. Obviously, the detection rate of such a technique is not satisfactory when the noise ratio is low. Moreover, if the variances of isolated regions in an image are equal, then the method fails to detect the forgery. In essence, the existing methods of splicing detection have high accuracy rate but most of the methods have time complexity issues and validity for geometric transformations is not that satisfactory. 3. Proposed work In order to detect image forgery, we propose a detection method that is based on the DWT and DCT. First, the lowest frequency sub-band or approximate coefficients are extracted by applying a DWT to the image. This is followed by the block-wise application of DCT to get the energy rich few coefficients. It will not be out of place here to briefly describe the motivation to use both DWT and DCT. Our approach relies on the division of the image to overlapping blocks for comparison, with a sliding factor of one pixel. Hence, in the absence of any padding, for a M × N image, there will be a total of (M − m + 1 ) × (N − m + 1 ) blocks of size m × m. This may lead to a huge number of comparisons, as each block must be compared with every other block. The situation is exacerbated if the image is large. As most of the information in an image is redundant, we believe the better way is to reduce the size of the image without losing most of the important information. The best way do so is to pass the image to frequency domain which would concentrate most of the energy in a few low frequency coefficients. At the same time, we do not want to lose the localization information, as the image has to be subsequently divided to overlapping blocks. With this in view, DWT was chosen as the transform to reduce the image size without compromising the localization information. DWT performs far better than DCT in getting an approximation of the image, especially in the absence of blocks. In the subsequent step, DCT is used because we are dealing with small blocks and for small blocks DCT is a better choice to compact the information. Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
ARTICLE IN PRESS
JID: CAEE 4
[m3Gsc;April 5, 2017;5:46]
K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
Fig. 1. Block diagram representation of proposed Forgery detection method.
Step by step process of the proposed detection method (Fig. 1) is as follows. 1. Consider gray scale image as input, if the image in RGB format convert to YCbCr and use Y as gray-scale image. The image is then subjected to necessary pre-processing, like noise removal/Gaussian smoothing and normalization. 2. The preprocessed image is then subjected to level l DWT to get 3l + 1 sub-bands. The value of l is proportional to the dimensions of the image. The larger the image, larger should be l and vice versa. However, the value of l seldom exceeds 5 and that too for very large images. The lowest frequency (high energy) sub-band LLl is chosen for further processing. The reduction in feature set size is dyadic and for a square image of size n × n, the lowest frequency sub-band has a size in the order of nl × nl . In other words, a level-3 DWT will give you approximate coefficients amounting to 1.56% of 2 2 the total. 3. The approximate lowest energy sub-band from the last step is divided into fixed size overlapping blocks. Although the DWT may have considerably reduced the number of coefficients, the number may still be large in face of the fact that we are dealing with overlapping blocks. If we continue the dyadic example from the preceding step, a sub-band of size n × nl may result in ( nl − m + 1 ) × ( nl − m + 1 ) square blocks of m × m coefficients, where m is the block size. Hence l 2
4.
5.
6.
7.
2
2
2
total number of coefficients, to deal with, are still ( nl − m + 1 )2 × m2 , which is a large set. 2 DCT is applied to each individual block in order to get the most energetic of coefficients from each block. Extract the most important (read energetic) features from each block in the form of a row vector, by following a zigzag order while starting from the DC coefficient. Note that we have now as many row vectors as were the number of blocks. Sort the block features, lexicographically, so that similar features become closer to each other. Note that lexicographic sorting is similar to that of dictionary, except with the additional criteria of size; sort by size first and then among the same size sort like a dictionary. (Fig. 2) Find the correlation coefficients between all the row vectors. Note that for our example from Step 3, we had k = ( nl − 2 m + 1 ) × ( nl − m + 1 ) blocks converted to as many vectors in step 4. Since each block i need to be compared with 2 each block j (j = i) to get the correlation coefficients of the form ρ ij , we end up with a total of k(k − 1 )/2 correlation coefficients. Note that −1 ≤ ρi j ≤ 1. Finally map the duplicated blocks by comparing all the ρ ij to a positive threshold ρ 0 . If for two blocks indexed i and j, |ρ ij | < ρ 0 , then mark them as duplicated. The marked blocks, if any, are indicators of the copy/move forgery.
It can be seen that the purpose of DWT is to reduce the image features to manageable size and then divide the latter to block that are made manageable by DCT application. Even then the number of coefficients may become very large if block sizes are not chosen with care. Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE
ARTICLE IN PRESS K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
[m3Gsc;April 5, 2017;5:46] 5
Fig. 2. Block diagram representation of image forgery by mask method.
4. Experimental results The proposed technique has been implemented and tested with images of various sizes and duplication at different regions therein. We have designed our experiments according to the following points: •
• •
Forge the image intelligently either using a state of the art image editing tool or by using the mask based method developed specially for this work. Apply the proposed detection method to the forged images. Compare the ensued results with those obtained by the application of forgery detection methods outlined in [10,24].
4.1. Image forgery by using the mask based method Alongside the photo-editing tool, we have developed our own editing method for a better idea on forgery detection through our method in comparison to others. The method consists of developing a binary multiplication mask in order to extract the part to be substituted as forgery in the original image. This method is applicable to copy/move forgery. The following steps are involved to forge the image: • • • •
Load the original RGB image, I1 . Select the region to be moved or copied. By using some image editing tool extract the specified image region that is to be placed. Convert this extracted region into a gray scale image.
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE 6
ARTICLE IN PRESS
[m3Gsc;April 5, 2017;5:46]
K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
Fig. 3. Image representation of image forgery by mask method.
•
•
Employ an edge detection technique, like Canny edge detection2 algorithm, in order to detect edges in the extracted region. Morphological processing is performed in order to complete the following tasks: 1. Dilate image to link the broken edges. 2. Hole/region filling
•
• •
The result would be a binary mask. Extract the object of interest by multiplying with the original, I1 . The result is an image (I2 ) having the object of interest with the rest being the background with zero intensity values. One can stretch the background to the size of the image to be forged (call it I3 - it is possible to have I1 = I3 ). In the correspondence between I3 and I2 , substitute a pixel of I3 with I2 , if the intensity of the latter is non-zero. Return the resultant forged I3 image with, apparently, no evidence of tampering.
Fig. 33 demonstrate the use of the mask based method for region duplication. We have tampered several images by using the above method and the image tampering results appeared seamless. The method is also capable to successfully forge via image splicing. Fig. 4 illustrates two more examples of forgeries4 . 4.2. Methods from literature used for comparison With regard to similarity to our method, we had chosen two methods for comparison to ours, in order to judge the effectiveness of the latter. The first method, by Cao et al. [10], will be hereafter called Cao’s method. The second method is by Zimba et al. [24] and will be hereinafter called Zimba’s method. 4.2.1. Cao’s method [10] This method is claimed to be robust against the detection of wrong similar blocks, having lower of feature vector dimensions and reduced time complexity. Step by step process of this detection method is as follows: 1. 2. 3. 4. 5.
The suspect RGB image is divided into fixed sized block. DCT is applied, followed by quantization, to get quantized coefficients. The blocks are represented by the circle block to extract fewer features in order to reduce the dimensions. Check the similar block pairs by correlation coefficients. Find the correct blocks and make a color map of duplicated block
4.2.2. Zimba’s method [24] This method detects image forgery by using a DWT to get the low frequency sub-band of the investigated image and then PCA is used to reduce the feature dimensions. Specifically, this method is used for the detection of copy/move image forgery or cloning. This algorithm consists of the following steps. 2 3 4
http://www.cse.iitd.ernet.in/∼pkalra/csl783/canny.pdf. Example image taken from www.hdwpics.com. Example 2 is after www.dreamstime.com and Example 3 is after www.hdwallpapers360.com.
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
ARTICLE IN PRESS
JID: CAEE
[m3Gsc;April 5, 2017;5:46]
K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
7
Fig. 4. Example forgeries. Table 1 Specification of detection algorithms.
1. 2. 3. 4. 5. 6. 7. 8.
Detection algorithm
Extraction domain
Block size
Feature dimension
Proposed method Cao’s method Zimba’s method
DWT-DCT Block based DCT DWT-PCA
8× 8 8× 8 8× 8
10 4 7
Consider a gray scale image, if the image is in RGB format then apply transform on each component separately. Apply the Wavelet transform to get the image details. Low frequency band is selected for further processing. Divide the image into overlapping blocks. For dimension reduction PCA â;; Eigen-value Decomposition is applied to get the row vectors. Row vectors are sorted lexicographically; as a result the similar vectors closest to each other. Compute the normalized shift vector between the adjacent block positions. Map the result on the basis of a preset threshold value.
Applicability of any copy/move forgery detection algorithm depends on the requirement that it should detect the duplicated regions which are either natural or forged. The specification of detection algorithms are shown in Table 1. 4.3. Experimental results of forgery detection The effectiveness of the forgery detection methods are usually gauged by the two measures given illustrated in the following equations:
r =| R
D|/|R|
w =| F − D | / | R |,
(1) (2)
where r is detection rate and w is false detection rate. R represents the actual copy/move tampered area and D is the detected area. Fig. 5 illustrates the detection rate accuracy of the proposed method in comparison to the reference methods for a number of tampered images from the testing set. The proposed method is superior in the sense that its average accuracy Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE 8
ARTICLE IN PRESS
[m3Gsc;April 5, 2017;5:46]
K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
Fig. 5. Detection rate accuracy of forged test images with different methods.
Table 2 Comparison table. Method
F.D.R
S.D
Detection accuracy
S.D
Proposed work Cao’s method Zimba’s method
6.47% 16.27% 10.31%
5.09 8.69 8.11
73.62% 72.77% 69.48%
10.74 9.85 10.23
is 73.62%, as against 72.77% and 69.48% for Cao’s and Zimba’s methods, respectively. Whilst the maximum accuracy for the proposed method (94.74%) is at par with Zimba’s method, it is far less for Cao’s method (87.88%). A better measure, to explain the spread, is standard deviation which was observed to be 10.74 as against 11.26 for Zimba’s method. The other method has a slightly better value of 9.85 but it suffers from lower average and maximum. Detection results of the proposed copy/move forgery algorithm are listed in Table 2, along with the results of the reference methods. The superiority of the proposed method is quite glaring with respect to the false detection rates (F.D.Rs), as can be observed from the graph in Fig. 6. With it’s high F.D.R., the Cao’s method is a poor performer; a fact also evident from its very high average F.D.R. – 16.27% as against 6.47% for the proposed. Zimba’s method is comparable to ours but still it’s average F.D.R is almost twice (10.31%). In fact the very low F.D.R. is the strength of our method, if read in combination with the standard deviation or S.D(5.09). For the two reference methods, the value is well above 8.
5. Conclusion The melange of DCT and DWT helped us achieving good results par rapport the two reference methods. The false detection rate was observed to be far less in comparison to the methods relying on either of the DCT and DWT. The low standard deviation of the false detection rate is like an icing on the cake. As of detection accuracy, the proposed method outperformed others by having highest average accuracy. The special multiplication mask based method, developed as part of experimentation, may serve in itself a seamless editing method. As can be observed from the example illustrations, the tampering is hard to be blindly noticed, i.e. only the presence of original image would establish the tampering. One additional advantage of the detection method was its viability for both copy/move and splicing based forgeries. All the aforementioned notwithstanding, there is a chance that the method may somehow under perform in the presence of occlusion. Same may be true of images with repeated patterns. In future, we intend to deal with these situations. We also plan to come up with a method to cover most of the contemporary forgeries. Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE
ARTICLE IN PRESS K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
[m3Gsc;April 5, 2017;5:46] 9
Fig. 6. False detection rate of forged test images with different methods.
References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26]
Mahdian B, Saic S. Blind methods for detecting image fakery. Aerosp Electr Syst Mag IEEE 2010;25(4):18–24. Shivakumar BL, Baboo SS. Detecting copy-move forgery in digital images: a survey and analysis of current methods. Glob J Comput Sci Technol 2010;10. Farid H. Digital doctoring: how to tell the real from the fake. Significance 2006;3(4):162–6. Zhang Z, Ren Y, Ping XJ, He ZY, Zhang SZ. A survey on passive-blind image forgery by doctor method detection. In: Proc. Seventh International Conference on Machine Learning and Cybernetics (ICMLC), Kunming, China; 2008. p. 3463–7. Lin G, Chang M, Chen Y. A passive-blind forgery detection scheme based on content-Adaptive quantization table estimation. Circ Syst Video Technol IEEE Trans 2011;21(4):421–34. Qazi T, Hayat K, Khan S, Madani S, Khan I, Kolodziej J, et al. Survey on blind image forgery detection. IET Image Proc 2013;7(7):660–70. Singhal N, Gandhani S. Analysis of copy-move forgery image forensics: a review. Int J Sig Process Image Process Pattern Recognit 2015;8(7):265–72. Qureshi MA, Deriche M. A bibliography of pixel-based blind image forgery detection techniques. Sig Process Image Commun 2015;39, Part A:46–74. Zhu Y, Ng TT, Shen X, Wen B. Revisiting copy-move forgery detection by considering realistic image with similar but genuine objects. CoRR 2016;abs/1601.07262. Cao Y, Gao T, Fan L, Yang Q. A robust detection algorithm for copy-Move forgery in digital images. Forensic Sci Int 2011. Fridrich J, Soukal D, Lukás˘ J. Detection of copy-move forgery in digital images. In: Proc. Digital Forensic Research Workshop; 2003. Singhal N, Gandhani S. Copy move forgery detection in contrast variant environment using binary DCT vectors. Int J Image Graph Sig Process 2015;6:38–44. Mahmood T, Nawaz T, Irtaza A, Ashraf R, Shah M, Mahmood MT. Copy move forgery detectiontechnique for forensic analysis in digital images. Math Probl Eng 2016;2016(8713202):13. Boz A, Bilge HS. Copy-move image forgery detection based on LBP and DCT. In: Proc. 24th Signal Processing and Communication Application Conference (SIU); 2016. p. 561–4. Warbhe AD, Dharaskar RV, Thakare VM. A scaling robust copy-paste tampering detection for digital image forensics. Procedia Comput Sci 2016;79:458–65. Kaur A, Shirma R. Copy-move forgery detection using DCT and SIFT. Int J Comput Appl 2013;70(7):30–4. Kaushik R, Bajaj RK, Mathew J. On image forgery detection using two dimensional discrete cosine transform and statistical moments. Procedia Comput Sci 2015;70:130–6. Myna AN, Venkateshmurthy MG, Patil CG. Detection of region duplication forgery in digital images using wavelets and log-polar mapping. In: Proc. International Conference on Computational Intelligence and Multimedia Applications (ICCIMA) - Volume 03; 2007. Zhang J, Feng Z, Su Y. A new approach for detecting copy-move forgery in digital images. In: Proc. 11th IEEE Singapore International Conference on Communication Systems (ICCS); 2008. p. 362–6. Peng F, Nie Y, Long M. A complete passive blind image copy-Move forensics scheme based on compound statistics features. Forensic Sci Int 2011. Khan S, Kulkarni A. An efficient method for detection of copy- Move forgery using discrete wavelet transform. Int J Comput Sci Eng (IJCSE) , 2010;05(02). Khan S, Kulkarni A. Detection of copy-move forgery using multiresolution characteristic of discrete wavelet transform. In: Proc. International Conference & Workshop on Emerging Trends in Technology (ICWET’11). ICWET ’11; 2011. p. 127–31. ISBN 978-1-4503-0449-8 Muhammad G, Bebis G. Blind copy move image forgery detection using dyadic undecimated wavelet transform. In: Proc. International Conference on Signal and Image Processing (ICSIP); 2010. Zimba M, Xingming S. DWT-PCA (EVD) based copy-move image forgery detection. Int J Digit Content Technol Appl (JDCTA) 2011;5(1):251–8. Liu Q., Sung H.A.. A new approach for JPEG resize and image splicing detection. In: Proc. First ACM workshop on Multimedia in forensics (MiFor’09). El-Alfy E-SM, Qureshi MA. Combining spatial and DCT based markov features for enhanced blind detection of image splicing. Pattern Anal Appl 2015;18(3):713–23.
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE 10
ARTICLE IN PRESS
[m3Gsc;April 5, 2017;5:46]
K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
[27] Sun S, Wu Q, Li G. Detection of image compositing based on a statistical model for natural images. Acta Autom Sin 2010;35:1564–7. [28] Jing T, Peng Y, Zhang F, Huo Y. Blind detection of digital forgeries using detection trace of eclosion. In: Proc. International Conference on Computational Intelligence and Software Engineering (CiSE’09); 2009. [29] Ying C, Yuping W. Exposing digital forgeries by detecting traces of smoothing. In: Proc. 9th International Conference for Young Computer Scientists (ICYCS); 2008. p. 1440–5. [30] Mahdian B, Saic S. Using noise inconsistencies for blind image forensics. Image Vis Comput 2009;27(10):1497–503. Special Section: Computer Vision Methods for Ambient Intelligence
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013
JID: CAEE
ARTICLE IN PRESS K. Hayat, T. Qazi / Computers and Electrical Engineering 000 (2017) 1–11
[m3Gsc;April 5, 2017;5:46] 11
Khizar Hayat is heading the Computer Science Section of College of Arts and Sciences, University of Nizwa, Oman. He has also led the Computer Science Department of COMSATS Institute of Information Technology, Abbottabad, Pakistan. He received his PhD degree in 2009 from the University of Montpellier, France, while working at LIRMM. His preference areas are image processing and information hiding. Tanzeela Qazi received her MS degree in Computer Science from the COMSATS Institute of Information Technology (CIIT), Pakistan in 2011. She mainly did her thesis work at the Abbottabad Campus of CIIT. Her area of interest is digital image processing with special reference to image forgery detection.
Please cite this article as: K. Hayat, T. Qazi, Forgery detection in digital images via discrete wavelet and discrete cosine transforms, Computers and Electrical Engineering (2017), http://dx.doi.org/10.1016/j.compeleceng.2017.03.013