Side-Informed Image Watermarking Scheme Based on Dither Modulation in the Frequency Domain

Digital communication and media sharing have extensively increased in the last couple of years. Research is focused on protecting digital media through copyright protection. The digital media is secured by watermarking. We have developed an image watermarking technique in the frequency domain to hide secure information in the Discrete Cosine Transform (DCT) coefficients of the carrier image. DCT coefficients are modulated by Dither Modulation (DM). We have increased the modulation step to be able to entirely recover the embedded data (watermark) and increase the robustness of our proposed algorithm to affine transforms and geometrical attacks. Our algorithm showed lower complexity and robustness against different attacks.


INTRODUCTION
Steganography is the art of hiding data into digital carriers.Steganography is used mainly in secure communication.Watermarking is a small branch of steganography.The watermark is embedded into the digital carrier to make it robust to different types of attacks.Even if the watermark is visible to some algorithms, it must be robust to the degree that makes it impossible for the attacker to use the digital carrier after successfully removing the watermark.
Watermarking research started in 1990 [1], watermarking is mainly used for copyright protection such as protecting governmental documents [2].The massive production of communication technologies and the insecurity of transferring digital data in digital channels made it urgent to find a method to protect digital media and secure digital communication.Generally, any watermarking technique must have the following properties: • Imperceptibility: This means embedding the secure information into the digital channel without detection from Human Visual System (HVS) or statistical methods.
• Robustness: This means the watermarking scheme must be robust mainly to affine transforms such as rotation, translation and scaling.In addition, it must be robust to geometrical attacks such as compression and noise addition.
Spatial domain watermarking techniques such as Least Significant Bit (LSB) are easily detected by Steganalysis tools [3].Extensive approaches in literature concentrate on embedding the watermark in the frequency domain.Li and Wang [4] presented watermarking approach that modifies the quantization table of JPEG compression and inserts the secure data into the middle frequency coefficients.They paid *Address correspondence to this author at the City for Scientific Research and Technology Applications, IRI, Alexandria, Egypt, Tel: +20125966704; Fax: +20125966704; Email: mohandesosama@yahoo.com attention to increases the payload; their new version of quantization table provides 36 coefficients in each 8x8 block.Watermarking based on DCT (Discrete Cosine Transform)of JPEG compression contains several stages, JPEG compression is shown in (Fig. 1).Watermarking is done in the quantization step to avoid losing the secure data.
In [5], the authors embedded the watermark by applying dual domain transforms into the carrier image.However, such techniques increased capacity but are still more complex to be implemented.Watermarking scheme based on Scalar Costa Scheme (SCS) is presented in [6].The author changed the random Codebooks of the Costa scheme which is presented in [7] with lattice-structured Codebooks in order to reduce complexity.The author also implemented a scheme independent of the host signal distribution.The author supposed independent and identically distributed(IID) signal and considered only the White-Gaussian-Noise (WGN) attack so that the scheme can be measured only by using Window to Noise Ratio(WNR).
In this paper, we proposed watermarking technique based on Quantization Index Modulation (QIM).QIM is a powerful watermarking technique; attention is paid to QIM as a powerful watermarking technique and adopted by researchers in the past ten years for its simplicity [8].Dither modulation (DM) -is a type of QIM, it will be used for encoding the watermark.Our approach showed simplicity in implementation (The algorithm is presented in a couple lines of code), it showed also robustness to affine and geometrical attacks.
The paper is organized as follows: in section 2, the proposed watermarking technique will be presented.In section 3, the experimental results will be presented to show the performance of the proposed algorithm.

THE PROPOSED WATERMARKING TECHNIQUE
The overall idea of the proposed watermarking solution is to extract a set of features from the image.These features are then modulated by DM in the transform domain.The DCT linear transform will be used to extract the coefficients of the image.The description of the two-dimensional DCT [9] for the input image F and output image T is calculated by: (1) Where And Where M, N are the dimensions of the input image, m,n are the values ranging from 0 to M-1 and 0 to N-1 respectively.
DCT coefficients will be modulated by using Dither modulation.The watermarked image will be obtained by getting the inverse of the modulated coefficients.To extract the watermark, we extract the positions of the carrier features from the stego-image.These features are obtained by applying DCT on the watermarked image, the positions which hold the watermark bits will be scanned until the length of the watermark is reached.
The decoding and encoding phases of the proposed algorithm are depicted in (Fig. 2).In the next subsections we are going to introduce the encoding and the decoding phases in detail.

The Function for Extracting the Image Features
There are many transforms used to extract the image features in the frequency domain.The most common transform is the DCT linear transform.The main objective of the proposed approach is to slightly change the DCT coefficients by using one of the side-informed image watermarking techniques.Because of the slight change, perturbation will be noticed in the image after watermarking.However, watermarking in the frequency domain is more robust than the traditional spatial domain embedding methods such as LSB.Our feature extraction procedure is going to sort the coefficients in descending order.The feature vector contains a number of coefficients equal to the size of the watermark.

Watermark Encoding by Using a Side-Informed Watermarking Scheme
QIM (Quantization Index Modulation) is selected as a side-informed watermarking scheme.The main idea of QIM is to move x to the centroid of the nearest decision region.In case of noise addition, the embedded watermark will not be affected since the change will sway x around the center of the region, probably x will not move to another region after attack.Dither modulation is the shifted version of the basic quantizer.
• If c is the centroid for logic 1 then c + d is the centroid of logic 0.
The basic quantizer will be uniform with quantization step S = 2d and d is the dither value.
The watermark encoder is depicted in Algorithm 1.

Read the watermark Calculate watermark size wmsz Read the host image Assign quantization step S with suitable value (application dependent) Get DCT of the host image Extract X vector containing the largest coefficients, number of f =wmsz
For each feature f in X Get the next bit to be embedded Wi Calculate the integer quotient Q and the remainder R in the following way [10] Where % is the modulo operator Make Q%S =Wi Always hold by applying the following equation Where f' is then new value of the feature.end for Get the watermarked image by DCT inverse.
So, to embed one bit value Wi, we shifted the centroid position of the feature f so that Q is an even value for the bit value 0, or an odd value for 1.

Watermark Decoding
The decoding process starts by taking the DCT of the watermarked image.The DCT of the watermarked image will be explored to extract the watermark bits.The decoding algorithm is depicted in Algorithm 2.

Read the watermarked image Get DCT of the watermarked image Extract X vector containing the coefficients which carried the watermark bits
For each feature f in X Get the next bit by the following equation Wi=Q%2 end for

EXPERIMENTAL RESULTS AND DISCUSSION
The experiments and results are tested on the University of California image database [11].A description of the dataset will be introduced then we will explore the effect of Host to Watermark Ratio (HWR) on the quality of the image.JPEG compression affects the decoding process and decreases the ability of extracting the watermark.JPEG compression will be plotted with respect to WNR and with different capacity.Finally, our algorithm will be tested for robustness against different attacks.

Dataset Description
We have used a set of images downloaded from the University of Southern California [11].The miscellaneous volume of the database consists of 44 images, 16 colors and 28 monochromes.The sizes are fourteen images of size 256x256, twenty six images of size 512x512, and four images of size 1024x1024.We have implemented our algorithm by using MATLAB R2009a, in the algorithm we have resized all the datasets to be 512x512; and for colored images we converted them from RGB to gray-scale images.Each experiment is applied on all the datasets.We have displayed the average results for each experiment.So the curves, tables and figures are not showing the actual result of specific image, they show the average result after applying the algorithm on all images in the dataset.

The Effect of HWR on the Visibility of the Watermark
Two power ratios are important in the context of watermark embedding [12].The HWR is a measure for the embedding power of the watermark: ( It should be kept as high as possible to guarantee a negligible perceptibility of the watermark w.The WNR characterizes the watermark channel quality (3) We have applied the above function on samples of the dataset after watermarking them with the watermark in (Fig. 2), we found HWR =31 dB.
In Fig. (4) if HWR is increased, it will have an effect on increasing the visibility of the watermark in the host image.i.e. decreasing the quality of the host image.So, it is approximately linear relationship.

The Effect of JPEG Compression with Different Quality Factors on the Watermark
JPEG compression affects the quality of the decoding process.In Fig. (5a), for quality factors between 5 and 80 there is little change in WNR.For values more than 80, WNR rapidly increases.All analyse are applied on the images with HWR=15 dB.If we consider capacity in our experiment, as shown in (Fig. 5b), we notice that, Increasing bit capacity degrades the extracted watermark.For high quality factor and low capacity, the decoding process has less probability of error.However, for high capacity and low JPEG quality factor the extracted watermark is extensively degraded.

Robustness of the Proposed Encoding Technique Against Different Attacks
Our proposed watermarking scheme is robust to affine transforms such as, rotation, scaling and translation.With  affine transforms, the image intensity is not changed and hence DCT and inverse DCT feature extraction will not be affected.The only thing we need is to locate the carrier coefficients; this is done in one of two ways, either to make the watermarking non-blind and use the original image in the decoding process, or use extra information embedded in the image about the location of the carrier coefficients.Peak Signal to Noise Ratio (PSNR) is used as a measurement of the image quality after different attacks.
PSNR can be defined by using the mean square error (MSE).Given m n monochrome image I and the image after attack Is, MSE is defined as (4) The PSNR is defined as: (5) Here, 255 is the maximum possible pixel value of the image represented using 8 bits per sample PSNR can be represented in the compact form as: (6) As shown in Table 1, for scaling attack, the watermark is not affected at all.For cropping attack, the algorithm is not robust, the image intensity will be changed and all its components and features will be different in the DCT and inverse DCT feature extraction.The locations of the carrier features are lost and the quality of the image is also affected.For JPEG compression with quality factor 80% (expressed in the table as attack=0.2or 20%), most components of the image features are not affected and the decoding process can successfully recover more than 99% of the watermark.As JPEG compression quality decreased the recovered watermark is negatively affected, but still the recovered logo can be distinguished and identified.For noise addition such as adding Salt & Pepper noise and Gaussian noise, the watermarking algorithm is robust, more than 80% of the watermark can be recovered and the watermark can be easily identified.

Performance Comparison of the Proposed Technique
The authors in [5,8] proposed a similar technique but with increased levels of embedding.However, they increased the complexity of their approach by watermarking in dual transform domain.We have reduced the complexity of the algorithm by embedding only in DCT domain.The execution time is reduced compared to previous approaches.
Table 2 lists the comparison between the selected previous approaches and the proposed approach.The algorithms
Fig. (3) shows the DCT transform of the Lena image.The high frequencies are located in the lower right corner of the image, changing them will dramatically change the quality of the watermarked image.The low frequencies are located in the upper left corner of the image, changing them will not affect the watermarked image.

Fig. ( 4 ).
Fig. (4).Visibility of the watermark in the host image as a function of HWR (a) The image with embedded watermark HWR=31 db (b) The general trend of the changing of HWR with respect to image quality.