A Robust Color Image Watermarking Scheme using Chaos for Copyright Protection

An exponential growth in multimedia applications has led to fast adoption of digital watermarking phenomena to protect the copyright information and authentication of digital contents. A novel spatial domain symmetric color image robust watermarking scheme based on chaos is presented in this research. The watermark is generated using chaotic logistic map and optimized to improve inherent properties and to achieve robustness. The embedding is performed at 3 LSBs (Least Significant Bits) of all the three color components of the host image. The sensitivity of the chaotic watermark along with redundant embedding approach makes the entire watermarking scheme highly robust, secure and imperceptible. In this paper, various image quality analysis metrics such as homogeneity, contrast, entropy, PSNR (Peak Signal to Noise Ratio), UIQI (Universal Image Quality Index) and SSIM (Structural Similarity Index Measures) are measures to analyze proposed scheme. The proposed technique shows superior results against UIQI. Further, the watermark image with proposed scheme is tested against various image-processing attacks. The robustness of watermarked image against attacks such as cropping, filtering, adding random noises and JPEG compression, rotation, blurring, darken etc. is analyzed. The Proposed scheme shows strong results that are justified in this paper. The proposed scheme is symmetric; therefore, reversible process at extraction entails successful extraction of embedded watermark.


INTRODUCTION
of digital contents over insecure channel i.e. internet.
Watermarking seems favorable for protecting the intellectual property rights of the multimedia contents from malicious attackers [1][2].
Digital image watermarking is serves as a method for information hiding. A watermark, which is company logo, producer name, social security number etc. is embedded into the host image. A good watermarking scheme cannot degrade the quality of the media. Further, it makes reliable retrieval of the embedded watermark for the authorized entities. An efficient watermarking method should keep the balance among imperceptibility, security, robustness and capacity [3][4].
A great amount of research has been done to design spatial domain watermarking schemes [5][6][7][8][9]. It is because to achieve the requirement of high embedding capacity and low computational complexity. Typically, the least significant bitplane of host image pixels is modified in spatial domain watermarking schemes using arbitrary chosen watermark.
Thus, it improves the imperceptibility with reduced complexity [10][11][12]. Chaos is a nonlinear dynamical system that is favored to design spatialdomain watermarking schemes. It is favored due to its properties of sensitive dependence on initial conditions, mixing and ergodicity [13][14]. Chaotic maps are used to design simple and efficient watermark in terms of nonlinearity and complexity.
A number of techniques have been found to design chaotic watermarks [15][16][17]. A chaos based LSB watermarking scheme was proposed in [18], in which logistic map was used to generate watermark sequences that were redundantly embedded into the host image. During the course of embedding, host image was divided into two sets and modifying the relationship between these two sets performs watermark embedding. In [19], a color image watermarking algorithm was presented, which utilized blue component of host image for watermark embedding. Pixel values were modified by DC (Direct Current) coefficient distribution principle. Watermark embedding was performed four times into different locations of host image. The scheme achieved results in terms of easy implementation in spatial domain and high robustness in frequency domain. In another scheme [20], 4x4 S-Box was generated by applying affine transformation on elements of a given number in GF (2 4  A blind dual watermarking algorithm for color images was discussed in [7], where invisible robust watermarks were embedded into the YCbCr color space by using DWT (Discrete Wavelet Transform) and fragile watermarks were embedded into the RGB components by utilizing LSB method.
An SVD based scheme for blind watermarking is presented in [22]. Modifying the U matrix embeds the watermark.
Although permuting the watermark with Arnold cat map preserved the security and robustness of the certain regions of watermarked image, but when dealing with pixels of edges and textured regions, the scheme did not perform well. This happened because watermark embedding was restricted to specific regions of the image not into the whole image.
Further, another spatial domain blind watermark method is recently presented [23]. In this method DC coefficient of pixels are modified in spatial domain. In addition, four subwatermarks of main watermark are embedded into different location of plain image to achieve better invisibility and strong robustness.

METHODOLOGY OF PROPOSED SCHEME
The methodology of proposed color image watermarking scheme is presented in this section. In proposed scheme, chaos based highly dispersive watermark is efficiently embedded into the host image. The watermark is generated using chaotic logistic map with optimization assumptions so that the robustness and security of the watermarked image can be improved. To do so, S-box is generated that is employed as a watermark in this work.
For highly dispersive watermark, the design assumption is that, for a given input difference/prediction, a good Sbox entails distinct difference between positions. A design assumption is given as follows [24]: A design based on this assumption usually implies that, in case, a given S-box does not meet the criterion that the repetition of any output difference is minimum 2 for all input differences, one is forced to look for a completely new S-box. This S-box, on the other hand, proposes an incremental design technique, where the Sbox is built up incrementally. An incremental procedure would entail starting with some tentative initial S-box whose first two entries are first tested. These entries are retained, and next entry is tested for how many repetitions of each output difference are for every input difference. If this entry also meets the criteria that the repetition is not more than twice, the entry is retained.
Else one iterates the chaotic map to regenerate new entries.
The pseudo code for the proposed scheme is given in Appendix-1 just before reference section. In this work, redundant embedding approach is employed to disperse multiple copies of watermark into the host image. It is to preserve high embedding capacity and to achieve robustness in spatial domain. Watermark embedding and extraction processes are given in Fig. 1(a-b) respectively.
The detailed steps involved to design proposed scheme are mentioned below: Step-1: Initially take any color image "H" of size 512x512 as host image Step-2: Extract the RGB components of "H" such that: Step Where H c denotes the components of H.
The proposed algorithm employs 16x16 block size because the chosen watermark is also of size 16x16. Moreover, it is more economical to use smaller sized blocks i.e. 16x16 instead of bigger sized blocks i.e. 32x32 and 64x64 as they have direct impact on the computational complexity of the algorithm.

Image Textural Features Analysis
This section measure watermarked image using proposed scheme. In order to test the robustness and imperceptibility of proposed method, different statistical analysis are considered [26]. Table 1 reports the five security features obtained from GLCM (Gray Level Co-Occurrence Matrix) [27][28]. Entropy It denotes the spatial disorder, where p(x,y) denotes the number of GLCM.

Contrast
It denotes the local variations ratio in GLCM.

Correlation
It performs comparison between the two images by considering joint probability occurrence of specified pixel pairs. Where, x,y are the image pixel positions, Pxyis the number of is the variance and standard deviation.

Energy
It represents pixel repetition rate as well as uniformness of the texture in a particular distribution.

Homogeneity
Measure non-zero entries in GLCM Where x and y are the host and watermarked image, xy are the mean values, xy are the standard deviations an d xy is the covariance. Based on the above three equations, the UIQI metric can be represented as in Equation (20).
The SSIM metric measures the similarity between two images by using three components of luminance, contrast and image structure as given in UIQI. The mean, standard deviations and covariance of host and watermarked image are computed using Equations (21-23) respectively.
The overall structural similarity index can also be computed as in Equation (24).
DSSIM is measure distance derived from SSIM. This metric is used to measure the dissimilarities between the host and the watermarked images. Mathematically, it is easily measured using SSIM given as follows: The results of MSE, PSNR, RMSE, SSIM, UIQI, and DSSIM, for eight reference images of "Lena", "Peppers", "Baboon", "Airplane", "Tiffany", "Splash", "Lake" and "House" is given in

Imperceptibility Analysis
Both the PSNR and SSIM metrics are widely used to measure the watermark imperceptibility. The embedded watermark should remain invisible into the image and human eyes should not perceive its presence or absence.   (26) Where W ij is watermark, (i,j) are the watermark index and W ' ij is the value at index (i,j) of the extracted watermark. The NC values are in the range [0 1]. Generally, it has been observed that NC values greater than 75 means that the two images are highly correlated to one another. The proposed method is compared with method [7], method [21] and method [22] in terms of NC. BER is a measure of bit errors to the total number of transferred bits in specified time interval. Mathematically, it is expressed as:

A Robust Color Image Watermarking Scheme using Chaos for Copyright Protection
Where 1i M and 1 j N. W ij is embedded watermark and W ' ij is the extracted watermark. The salt and pepper noise attack with varying intensities of 0.01, 0.02, 0.04 and 0.08 is performed on watermarked image using proposed method. The proposed algorithm shows NC=1 in all the cases, which means perfect extraction of the embedded watermark. All other schemes for comparison such as Method [7], Method [22] and Method [21] have lower values of NC, which ensures the robustness of proposed algorithm against well-known attacks. The NC results for all the four intensities are given in Table 5 and the visual appearance of the attacked image with intensity 0.02 is given in Fig. 4(a). The NC values of the watermarked image against blurring and Gaussian noise attacks are reported in Table 5 Table 5 shows the NC results against these attacks and Fig. 4(d-e) show the results of NC in case of brighten 80 and darken 80 attacks.
The attacks of resizing, contrast and collage are applied to the watermarked image "Lena" and results are shown in Table 5. It is evident that after resizing 50%, the proposed method shows improved results as compared to another scheme. Further, in case of contrast, proposed method shows improved results as compared to method [21]. The results of proposed method against 30% collage attack compare with other techniques show strong robustness of proposed scheme. Fig. 4(f-h) Table 6. The strength of proposed algorithm is verified from the tabulated results that even after degrading the quality i.e. compressing image with higher compression level (Q=20), we are still able to achieve NC=0.8481 while operating in spatial domain.  Fig. 5 shows the comparison results of proposed method with method [29] and method [30] in terms of BER. The low values of BER in case of JPEG (Q=30) attack as compared to method [29] and method [30] proves the robustness of proposed method against this attack. Because real coefficients of middle frequency band are modified by a factor in method [29]. Thus watermark did not survive against higher compression levels.
It is observed that in method [30], DCT coefficients are modified to reach the difference between two adjacent blocks in particular range. Therefore, the modification performed through scaling operation would not resist against higher compression attack. Fig. 6 shows the result of NC and BER for watermarked "Lena" image after different percentages of cropping.
Along with NC and BER metrics, the proposed technique also shows the results of PSNR measure against cropping attack as given in Table 6.

FUTURE WORK
The future scope of this research work will include the usage of three different chaotic map-based watermarks embedded in two different domains.