A Robust Multiwatermarking Scheme for Multiple Digital Input Images in DWT Domain

Transcription

1 International Journal of Computer and Information echnology (IN: ) Volume 3 Issue 4, July 214 A Robust Multiwatermarking cheme for Multiple Digital Input Images in DW Domain Mahdi Babaei Faculty of Information & Communication echnology, imkokwing University of Creative echnology, Malaysia mahdi.babaei {at} limkokwing.edu.my Kok-Why Ng Faculty of Computing and Informatics, Multimedia University, Malaysia HosseinReza Babaei Faculty of Information & Communication echnology, imkokwing University of Creative echnology, Malaysia Hamid Ghorbani Niknajeh Faculty of Computing and Informatics, Multimedia University, Malaysia Abstract Digital ing is an encryption technique that protects the ownership of multimedia using algorithms. It has opened a new domain in multimedia security and protection against attacks. In this paper, an experiment is conducted using four inputs s. In order to propose a Multi-ing method using four inputs we have applied an inductive computation that improved the quality of output from one side; and robustness of trademarks which are extracted from the process on the other side. he process contains four two dimensional trademarks which will be converted and combined into a three dimensional sequence. he next phase is to use sub-digital s in Discrete Wavelet ransform domain and decompose them into non-overlapping blocks and finding the block which contains the most densely packed information pertaining to texture (based on the size of the binary ). he latest stage is to embed the final Multi- before adjusting the selected block pixels. his will be based on some discrete operation rules, which will later be discussed. he results will be compared to currently available three-input methods. Keywords - Digital ing; Wavelet ransform; Gaussian; Image filtering. I. INRODUCION he rapid development of communication technology and the prevalence of Internet as a worldwide network have facilitated the accessibility to information and media. With the advent of Multimedia-based information technologies, transmission and information sharing among social communities are made possible via audio, music and digitized videos. ince there was a huge demand on information protection and publishing copyright in the market, the information concealment techniques became very popular. It can be divided into various categories. teganography[2,3] and digital ing[4] are two main categories which are currently available in the market. he art and science of concealing information enable certain information to go undetected called teganography [1]. he Digital ing is a technique to secure digital contents by combining the original digital and a watermark [5]. It would be possible to transmit the watermarked and extract it on receiver side. s comprise of visible and invisible s. It can be invisible when the range of application is wide. he business logos for most small-sizes businesses use the visible watermark s to ensure copyright protection. In this process, the watermarked s will appear similar to the original s while having one or more hidden watermark s which can be extracted. he main usage of ing is for identification, content authentication and copyright protection. uch processes require a robust watermark algorithm toward various attacks[6]. hose attacks are mainly related to transformations such as translation, scaling and rotation. he ing techniques against those kinds of attacks can be categorized into Blind and Non-Blind ing. he Blind [7]ing is when the original is not needed for watermark recognition, and the non-blind is where the original is needed. he limitations of patial domain-based methods such as [8] or [9] (both teganography and ing applications), have not made them a favorable choice for a robust Multi-ing technique. hey depend a lot on earlier techniques, like cryptography and encryption, whereby their lower capacity and reduced robustness in affining maps and distorting signal processing in the majority of the spatial domain based scheme. One of the main limitations of the Discrete Fourier transform (DF) and Discrete Cosine ransform (DC) [1] is their limit in the amount of tampering and attacks they can withstand. he low capacity of hidden data that Wavelet schemes in [11] and [12] which it can sustain is considered the most important constraint under these methods. In addition, the scheme also lacks robustness with respect to geometric attacks

2 International Journal of Computer and Information echnology (IN: ) In addition to the above lines, all information hiding in the DW, DC, and DF of the carrier s consist of a number of computation and information coding. In order to have good quality stereo s, these methods use the mid to low frequency to embed the hidden data. As a result, it imposes a tradeoff between the size of the hidden data and the quality of the resulting watermarked s. In general, this is another constraint for some of available ing chemes, aside from their robustness limitation to geometric attacks. II. A. General Algorithms IERAURE REVIEW he is an authentication method based on logo(as company or person letterhead), name or even stamp containing lines and texts to show that a certain document has been certified by an authorized person [13]. ransforming from traditional to the digital way has revolutionized Information security techniques which are heavily based on digital technology. his technology embedded into different types of media such as audio, video and photos. he major attempt is to hide a security code quietly in the media and this can be extracted (or recovered) on the receiver side. In general a digital watermark system consists of two major parts [5]: I. Embedding the watermark and secret key into the cover which is basically the original (Figure1). Cover Image ecret Key Figure 1-General way to embed watermark II. he second step is to detect and extract the watermark with the given secret key and watermarked in hand. his part executes on the received side after data transmission (Figure2). ed Not ed Embedding Detection Figure 2- Detection and Recognition ed Image ed Image ecret Key B. Previous Works According to [14] the two major categories for digital ing techniques are as follows: I. Pseudo-Random Gaussian equence: his type of ing is more useful when there is a need for objective detection. It is using 1 and -1 to show the watermark. II. Binary Image or Grey cale Image s: his type of watermark is more useful when there is a need for subjective detection. he logo s can turn into a binary or grayscale watermark and used for ing scheme. Furthermore, an can be represented in two different domains: I. patial domain: he is characterized and shown in shape of frequencies. II. ransform domain: he represents by pixels. he segmentation of would be applied based on frequencies. here are several algorithms in this type of ing and the popular ones are shown below: A) Discrete Cosine ransform (DC): DC is a simple processing procedure. It is fast in transformation and the real output is saved into JPEG. here are two disadvantages with this type of ing: 1.) although they are more robust in comparison to with the patial Domain algorithms, they are less effective and secure against simple types of geometric attacks like rotation. 2.) heir weakness extends even to scaling. hese types of techniques can withstand face very simple attacks like pixel brightness or contrast modification and readjustment or blurring. hey are not easy to implement and the calculation and infrastructures needed for this type of ing have higher prices compared to others. In fact, they are not much resource efficient. here are two types of DC ing available: a) Global DC ing b) Block based DC ing. B) Discrete Wavelet ransform (DW): he main application of this tool is to make multi-resolution decomposition available for both spatial and frequency. he general idea of DW for Image Data hiding and ing is to split the cover and decompose it into four sub- graph quarters [15]. In Figure3, H1, HH1 and H1 show the best scale wavelet coefficients where the 1 is the coarse level coefficients. H1, HH1 and H1 are the detail subs where 1is considered as the approximation sub-. It is also possible for 1 to be decomposed further into H2, H2, HH2, and 2. he process of decomposition continues until the final scale is reached

3 International Journal of Computer and Information echnology (IN: ) Figure 3- DW segmentation In low frequency, the stability is enhanced unlike the processing attack of all other subs acquired in Discrete Wavelet ransform. In other words, the decomposition consists of one low frequency zone (at least) and three high frequency sub-graphs. In DW, each of three high frequency zones shows information about s edge, s texture and finally contour. In this type of ing, the energy of the is mostly in low frequency. he higher robustness can be achieved through low frequency coefficients. Cover Image Embedding Figure 4-DW ing C) Discrete Fourier ransform (DF): DF ing had been proposed because of its robustness in facing attacks like rotation, scaling or cropping, which are generally Known as geometric attacks. It is also very effective when faced by frequent attacks. In each Fourier equation, coefficients can be split into two components: 1) Phase and 2) Magnitude. Because of the two components, DF is considered as complex valued. he complex output of this technique is one of its disadvantages. he strongest part of a DF is its centered components with low frequency. Based on the experiments, geometric attacks do not change the amount of Phase coefficient. herefore, this type of ing techniques can bring robustness against translation, rotation, and scaling attacks. III. H2 1 H1 H2 HH2 caling Factor DW IW H1 HH1 Choose Haar wavelet in high frequency band ed Image PROPOED OUION he idea of developing a robust DW multi- ing scheme was released in 212 by Y. Zhou and W. Jin[15]. his technique proved that the DW-domain algorithm is robust, time efficient and reliable against attacks. he limitation of the developed model is the number of watermark s are only three. We set the research goal to improve their techniques and to see the effect of increasing the number of watermark s in final result. he hypothesis would be the same as [15] but using different number of watermark s. In this paper, we apply the same strategy using four s in order to compare the results and find out the capability of the algorithm in inductive calculations. In case if we use the results analysis for inductive calculus, we would have K number of algorithms with the same steps but with different outputs. he correctness of the first two members shows that it would be possible to prove that this strategy may work for any number of the same algorithm. et s say: K is rue for the robust DW algorithm using three input watermark s resulting in a correct output. K+1 is rue for the same robust DW algorithm as K using four input watermark s and to check for correctness. We can conclude that: K + n is rue for any number of input watermark s. (n is an integer starts from 3) he proposed method using four watermarks and the six steps in this algorithm are shown as Figure5. EP 1 EP 2 EP 3 oad cover and apply Calculation of election of better the DW segmentation. blocks Uniformity. blocks. EP 4 Extraction of fractional parts. Figure 5- Proposed olution teps A. ing tep 1 he M x M-size is imported into the application. As this is in grayscale, the value of each pixel will be ranged from zero to 255. Based on the basic idea of DW technique, we split the cover into quarters. he -level approximation will be applied on the and we may say the size of each sub- would be M/ x M/. Cover size : M M F f i, j 255 M is size in pixel Image F size : M / 2 M / 2 F is the original grayscale f ( i, j) denotes original dimentions is approximation level EP 5 EP 6 Combine and Embed 4 s. generate the output. he size for each watermark (in our scenario four s) would be n x n and it will be an integer multiple of cover. We may change the multiple 2D watermark s into a 3D using the conditions below: 836

4 International Journal of Computer and Information echnology (IN: ) dimension: n n M / n r ( r int AND r 1) k xw w (i, j,1 : k) w (i, j, z)1 z k (1) ( i, j) tep 2 he cover has split into smaller s, which are called the sub-s. Each of them needs to be totally separated and there should not be any share pixels existing between them. In this case, the s are called as nonoverlapping sub-s. In next step, we proceed to calculate the uniformity of each non-overlap sub (block) using the equation below: 1 f i, j mk Block uniformity : d (B ) x (2) K 2 i, j BK n m 1 B is n n size block k m is the average value of Block k is weight correction (.6~. 7) he meaning of block uniformity in here is the uniqueness and visual capacity of an block. It means that a block with higher number of uniformity will have larger visual capacity. On the contrary, a block with lower visual capacity will be shorter in uniformity magnitude. he reason we take this step is to calculate the visual capacity of blocks in order to be used in next step. tep 3 he next is to select the blocks with large uniformity number. he selected blocks will be named as B. he refers to a non-negative integer which denotes the number of block(s=1, 2,, ). In addition to that each block has to satisfy a certain criteria to be selected. he requirement of selecting a block is the equality of pixel number of that particular block and the binary watermark that will be embedded. o we will have: B : selected block number he sum of pixel number in the selected blocks is equal to the embedded pixel number in watermarks. tep 4 he fractional part of each pixel value will be extracted from the selected clocks B in this step. At this point, the function floor(x) will be round off element x to the next integer towards minus infinity. Applying the formula below will give us the result of true pixels without fraction. herefore, we have: BF : DBF (i, j) BF (i, j) floor(bf (i, j)) (3) Floor(x):round to nearest integer toward minus tep 5 In this step, the binary watermark will be rooted into a selected field. he value for DB (i,j) which k denotes the value of that particular pixel minus its fractional part will be modified based on an algorithm shown below. In this algorithm, we have: g: logical table which is returned by the application. Dec(x): Conversion of a Binary vector to Decimal integer In this algorithm, the condition will be set based on the value of DB in that particular pixel. If the value is positive but less than.5, the value will be modified to.25 and watermarks would be the g output. If the value of DB is integer and is greater than.5, it will be modified to.75. In this case, the g output will be based on the remainder amount of division of Decimal to two. If DBF i, j.5, DBF i, j.25, lg i, j Dec w i, j,1 : k ; else end end i, j.75; i j Decw i j k DBF if mod Dec w i, j,1 : k, 2 lg,,,1 : 1; else lg i, j,,1 : 1; Decw i j k tep 6 In this step, the original integral part will be recombined with the modified fractional part of DB and outline the new fields B (i,j). he recombination of the pixels will give us the modified. In fact the replacement of the original old blocks B with the new blocks B (i,j) will help us to get a up-to-the-minute approximated sub- F. As a final point, the watermarked is the outcome of a new approximated sub- F and the original detail sub-s with executing -level inverse wavelet transform. We will have the below formula to explain what will happen in this step: BF ' i, j DBF i, j floor i, j (4) B. Extraction Based on this hypothesis that the position of watermark would be known from the potential of that particular approximated sub- we would have: tep 1 In the first step of decomposition of watermarked, we have approximated on the sub-. In this step, we test and select an _F with M M in the size. In this condition, we would apply a DW with - level, and find an approximated sub- _F. tep 2 In this step, we would split the watermarked into multiple blocks. It is important to make the approximation of the sub- _F on the way that none of the blocks overlap each other. he size of each block would be n n and a particular selected block would 837

5 Result International Journal of Computer and Information echnology (IN: ) be named as (s=1, 2 ). he selection of those certain blocks would be based on the position of the information recorded in the watermark embedded process. tep 3 After that, the decimal part of each pixel for the blocks will be extracted. his process may happen by subtracting the fractional part from the true value is each pixel of watermarked. his would help us to reach and extract the watermarks. he formula that we should apply for that would be: BF : DB F B F ( i, j) floor B F ( i, j) (s 1, 2,3,,) (5) tep 4 In final step, we will recover the multi-watermarks. In order to perform the recovery, we need to follow a discrete operation with refer to the fractional value of the selected blocks and the logical table g. It is worth mentioning that the function Bin (x) denotes the conversion of decimal integer (x) to a binary vector. EP 1 EP 2 EP 3 EP 4 Decomposition of watermarked plit the watermarked into multiple blocks Extract the fractional part of each pixel for blocks Figure 6- Extraction teps Recover the multiwatermarks IV. REU AND ANAYI An experiment had performed using four base-2 s with the size of 32 32(shown in figure 7), a cover (figure 8) and the developed application. he final result is a watermarked which is shown in Figure 8. f * i, j and f i, j and watermarked is the total pixel number of the embedded digital watermark. b error is the pixel values of the original is the total error pixel occurred in extracted watermark. Figure 8- Cover Image and Final ed (using four watermark s) he best comparison of results with the previous researches in area would be based on our method in compare to [15]. here are two series of experiments been done using each method and the results show less error rate in extraction and higher quality of watermarked s in proposed method overall est PNR watermarked (3 WM Images) PNR watermarked (4 WM Images) Figure 9 - Quality comparison of Proposed Method Vs. Previous works Figure 7- Four watermark Images In this application, the watermark s are named as and. he 8-bit cover would be a grayscale person with in size. In this simulation, the fourth level of wavelet decomposition and reconstruction is used for the person. he ratio of peak signal to noise (PNR) is a major factor to measure and determine the quality of watermarked. In this case we will have an error rate in extraction of watermarks. his rate will be evaluated by. he definition of these rates would be the formulas below based on[15]: 2 1 ME 1 db PNR 1log 6 *,, 2 i1 j1 ME 2 7 b n error b n f i j f i j N 8 PNR watermarked Error extracted logo (3 WM Images) PNR watermarked Error extracted logo (4 WM Images) able 1-ed Image Quality comparison he simulated results indicate that the proposed scheme has ideal imperceptibility and capacity, which are independent of the embedded ing number; especially the scheme can guarantee the embedded multi watermarks. Not only the error rate has reduced but also the quality of watermarked has increased

6 International Journal of Computer and Information echnology (IN: ) Re sul t Figure 1-Error rate in extraction of logo comparison A. Attack testing I- Gaussian Noise esting the reliability of the anticipated method alongside Gaussian noise attacks with mean of and variance of.1 are applied. he PNR of the original watermarked reduces to dB from dB with an error rate of 1.294%. he detailed simulated results under different intensity noise addition are presented in Figure 12, indicating the watermark robustness against noise attack. B. Attack testing II- Median filter With the purpose of testing the reliability of proposed method against the other types of attacks, a watermarked woman filtered by median filter by [7 7] in size window has been tested with low-pass filtering. In this case, there are lot of missing details and high amount of PNR been reduced in watermarked from db to dB. Figure 13 shows the detailed simulated results under median filtering with different window size. C. Attacks testing results and analysis Figure 12 and 13in below show the result of using Gaussian Noise and median filter. Gaussian Noise (Variance) PNR ed Noise = able 2-Result of Gaussian noise attack testing Error extracted logo (4 WM Images) Median Filter windows size PNR watermarked Error extracted logo (4 WM Images) Without median * * * * * able 3-Result of Median Filter attack testing Error extracted logo (3 case) Error extracted logo(4 case) D. Comparison of PNR in Gaussian noise attack he results of Gaussian noise attack had been shown in able 4 shows the improvement of PNR in our algorithm in compare to what had existed before. he results also show an increasing PNR ratio percentage by increasing Gaussian noise variance. Gaussian Noise (Variance) PNR ed (3 WM Images) PNR ed (4 WM Images) able 4-Comparison of PNR in Gaussian noise attack testing using 3 and 4 input watermarks 4 Re34.82 sul t Figure 11-Comparison of PNR in Gaussian noise attack testing using 3 and 4 input watermarks. V. CONCUION est Gaussian Noise (Variance) PNR edi mage(3case) PNR edi mage(4case) We have researched on four-binary input s embedded them together into a grayscale digital. he system is derived from the novel discrete operation. he application operation and simulation results signify that the anticipated scheme has ideal imperceptibility and capacity, which are not dependent on the embedded ing number, in particular, the scheme is able to undertake the embedded Multi-s, as well as, the one having protected from all kind of attacks. he results shown in ection IV can clearly certify that the proposed method can generate a robust Multi-ing method in DW domain and can even certify that the general processing attacks are not harmful to this ing technique. he results show that there is an improvement in quality in compare to [15], and the error rate in watermark extraction is reduced as well. he comparison of the results to earlier studies with three input watermark s shows that the robustness has increased. his result with increment of quality and error extraction rates has boosted the immunity against the attacks. he low rate of error and high 839

7 International Journal of Computer and Information echnology (IN: ) quality in reducing the attacks show that this can perform better than what had existed before. Based on what we have for the four input s, we can say that the results are well accepted. he truthfulness of K=4 watermark s and the previous works in this field (using K=3) can certify the hypothesis that this algorithm can effectively work for K=n (n is a non-negative integer). It seems that adding the number of input watermark s will reduce the error rate and increase the accuracy of the algorithm. As for future work, it can further be improved on its accuracy by using higher number of watermarks. However, this will consume a lot of processing power. REFERENCE [1] F. A. Petitcolas, R. J. Anderson, and M. G. Kuhn, "Information hiding-a survey," Proceedings of the IEEE, vol. 87, pp , [2] D. Artz, "Digital steganography: hiding data within data," internet computing, IEEE, vol. 5, pp. 75-8, 21. [3] A. Martin, G. apiro, and G. eroussi, "Is steganography natural?," Image Processing, IEEE ransactions on, vol. 14, pp , 25. [4] I. Cox, M. Miller, J. Bloom, and M. Miller, Digital watermarking: Morgan Kaufmann, 21. [5] M. Abdullatif, A. M. Zeki, J. Chebil, and.. Gunawan, "Properties of Digital Image ing." [6]. Elshoura and D. Megherbi, "Analysis of Noise ensitivity of chebichef and Zernike Moments with Application to Image ing," Journal of Visual Communication and Image Representation, 213. [7] I. Cox, M. Miller, J. Bloom, J. Fridrich, and. Kalker, Digital watermarking and steganography: Morgan Kaufmann, 27. [8] C.-H. Yang, C.-Y. Weng,.-J. Wang, and H.-M. un, "Adaptive data hiding in edge areas of s with spatial B domain systems," Information Forensics and ecurity, IEEE ransactions on, vol. 3, pp , 28. [9] W.-. ai, C.-M. Yeh, and C.-C. Chang, "Reversible data hiding based on histogram modification of pixel differences," Circuits and ystems for Video echnology, IEEE ransactions on, vol. 19, pp , 29. [1] K. olanki, O. Dabeer, U. Madhow, B. Manjunath, and. Chandrasekaran, "Robust -adaptive data hiding: Modeling, source coding and channel coding," in Proceedings of the Annual Allerton Conference on Communication Control and Computing, 23, pp [11] J. Xie, C. Yang, and D. Huang, "High Capacity Information Hiding Algorithm for DC Domain of Image," in Intelligent Information Hiding and Multimedia ignal Processing, 28. IIHMP'8 International Conference on, 28, pp [12] D. Zou, Y. Q. hi, Z. Ni, and W. u, "A semi-fragile lossless digital watermarking scheme based on integer wavelet transform," Circuits and ystems for Video echnology, IEEE ransactions on, vol. 16, pp , 26. [13].Paul. "What is Digital ing?" Available: 27 [14] V. M. Potdar,. Han, and E. Chang, "A survey of digital watermarking techniques," in Industrial Informatics, 25. INDIN' rd IEEE International Conference on, 25, pp [15] Y. Zhou and W. Jin, "A robust digital multi-watermarking scheme in the DW domain," in ystems and Informatics (ICAI), 212 International Conference on, 212, pp