The Embedding Performance of StegSVM Model in Image Steganography

ABSTRACT


INTRODUCTION
The concept of exchanging hidden information causes concerns in the field of information security [1]. Therefore, several methods have been applied for this purpose including cryptography, steganography, and watermarking to establish hidden communication. Secret data can be concealed in different cover media such as image, video, audio, and written text [2]. However, among all these methods, steganography has lately acquired more attention and becoming increasingly important in the field of computer security. It is a process that involves hiding messages in other messages [1] while watermarking is using to hide the secret messages in various medium [2]. The key difference between steganography and watermarking is that the goal of steganography is more to covert communication [3], which is why, the fundamental requirement of steganographic algorithm is that messages hidden inside any carrier file should not be sensible to human beings. The goal of watermarking on the other hand, is to ensure the hidden message always remains present in the digital media to provide solid proof of ownership [4], [5]. In this study, the focus is generally on steganography and specifically on image steganography. In fact, image steganography is the most popular and widely applied technique today [6][7][8]. Among the factors that led to this situation is because most business today are using digital images as the main medium to disseminate information to be more attractive such as in advertising and so on. There are two popular kinds of domain images that are normally used in image steganography, namely spatial and transform domains. In spatial domain the secret-message is directly embedded into the pixels of the cover-image, for example, by LSB substitution. Meanwhile, in transform domain the cover-image is first transformed into coefficients such as discrete cosine transform (DCT), discrete fourier transform (DFT) and discrete wavelet transform (DWT) domains [6], [8]. However, the most popular and extensively used in image steganography is DCT domain [3], [4], [7]. This is because DCT domain relatively consumes less power and has shown to be uncorrelated due to energy compaction into just one coefficient i.e. DC coefficient while compressing the coefficients values, and this makes it possible to construct relatively simple algorithms [7].
Most of images types have been used as the cover-image in these domains like Bitmap File Format (BMP), Joint Photographic Experts Group (JPEG) and Graphics Interchange Format (GIF) images with JPEG being the most common image format for Internet and local usage since it provides a large compression ratio and maintains high image quality [12][13][14]. According to Cheddad [7], image steganography is the field that remains untested and very few of companies and associations have published the requirements of the steganography algorithm evaluation. This is because the target of the steganography scheme evaluation should be clearly identified based on it purposes. As mentioned earlier, imperceptibility and robustness are the prominent criterias in evaluating the steganographic technique. Therefore, this paper proposes the model that has been developed is StegSVM in steganography. It evaluates this model in comparison with the robustness and impercecptibilty of the technique in steganography.

STEGSVM MODEL EVALUATION
Usually, Peak-Signal-to-Noise-Ratio (PSNR) is utilized to verify the perceptual transparency and fidelity of image steganography algorithms. It gives a measure of the statistical differences between a coverimage and stego-image. PSNR is good in providing qualitative rank order scores as long as the same content and the same algorithm are used [15]. The higher the PSNR value, the more effective the technique is. As such, the technique can be said effective if the PSNR value is more than 40dB. Cheddad et al. (4) have stated that if the PSNR value falls below 30dB, it indicates that the quality of the image is fairly low since the distortion caused by the embedding is noticeable. Consequently, a high quality stego-image should attempt for 40dB and above. Furthermore, a greater PSNR value means a lower degree of image distortion after the secret-message is embedded. In determining the degradation with respect to the host image, the researcher applies the PSNR metric (Peak Signal-to Noise Ratio) and MSE (Mean Square Error) to measure the distortion produced after the embedding process [16][17]. It is defined as: (1) where x and y are the image coordinates, P and Q are the dimensions of the image, S xy is the generated stegoimage and Cxy is the cover image, as shown in Figure 1. PSNR is often expressed on logarithmic scale in decibels (dB) [4]. There are two kinds of metrics that can be used in order to evaluate the robustness of secretmessage, the Normalized Cross-Correlation (NC) and Similarity Ratio (SR). Both of these metrics are described in the next sections, respectively.
The robustness of the secret-message can be evaluated using the Normalized Cross-Correlation (NC) between the original secret-message and the extracted secret-message. The NC is evaluated by varying the strength of each degradation process which is defined as: where, M = secret-message; M' = extracted secret-message In order to test the robustness, the Similarity Ratio (SR) can be calculated between original and watermarked images. It can be achieved by using the following equation: The number of matching pixel and another different pixel values are represented by S and D respectively. If the value of SR is closed to 1, its shows the robustness of watermark is better and preserved [22]. According to Tsai et al. [23] NC has been commonly used as a metric to evaluate the degree of similarity between two compared images because of these two advantages. It is less sensitive to linear changes in the amplitude of illumination in the two compared images that makes measurement more accurate ii) It is confined in the range between -1 and 1 that the setting of threshold value is much easier because it involves the calculation of a smaller number. It is well known that NC can be efficiently implemented in the transform domain rather than spatial domain [24]. The NC value can be easily evaluated by using NC application as shown in Figure 2 as follows [20].

Figure 2. Normalized Cross-Correlation
Therefore, PSNR and NC have been selected to be used in this research to measure the imperceptibility of cover-image and the robustness of secret-message respectively. For the PSNR, the ImageJ application will be used as a tool to measure the cover-image [21]. Its power and flexibility allow it to be used as a research tool by scientists in various disciplines [25], including image information hiding. Meanwhile, the NC Tool has been chosen to be used in order to find the NC values between the original and the extracted secret-message. It has been used because of the availability and also as a convenient sample.
The evaluation is based on the value between the StegSVM model with the existing SVM classification models or with the LSB substitutions techniques, or with kinds of image processing attacks. As mentioned earlier, imperceptibility and robustness are the two most important criteria [24,25,16] in evaluating an image steganography. The imperceptibility and robustness of StegSVM is based on the comparison between first, the proposed and existing model and second, between the proposed model and other models of LSB substitution. For both of these experiments, the secret-message size has been set to 1024bits while the penalty value of parameter C and gamma is set to 30 and 0.5 respectively.

RESULT AND DISCUSSION
The result of PSNR by comparing two types of LSB techniques: Shifted LSB and Direct LSB Substitution is shown in Table 1. Shifted LSB is applied with the usage of StegSVM while Direct LSB Substitution technique is represented by using another tool that is StegaMage. is not only succeeded in embedding randomizing implementation, but it also makes sure that the embedding location is secured compared to direct LSB substitution. As for robustness, the proposed model will be evaluated based on the comparison of the value of normalized cross-correlation (NC) for extracted secretmessage between first, the proposed and existing models. Second is between the proposed model and other models of LSB substitution. Third is the comparison between the results of image processing attacks on the proposed and existing models. As for these three experiments, the size for the secret message has been justified to 1024 bits while the value for penalty parameter C and gamma is 30 and 0.5 respectively. Then, in Table 2 the shifted LSB technique recorded higher NC value as representative in robustness between the proposed model and other model of LSB. Based on Table 2, the comparison proposed model with Direct LSB Substitution Technique with the value range from 0.98 to 1.0. Simultaneouly, Direct LSB Substitution Technique recorded NC value that is much lower, being the highest value is 0.95 and the lowest is 0.84. Lena and Baboon Cover-image recorded higher NC value for both techniques compared to other cover-image. It is found that the NC value is much higher in comparison between the proposed model, the existing models and different LSB substitution method. In Table 3, after a few image processing attacks, the extracted secret-message is valued using NC.

CONCLUSION
This paper performed the evaluation of the proposed model StegSVM model compared to the other model through imperceptibility and robustness of the technique. StegSVM model were acceptable in which it shows a higher quality steganography, thus enhancing the performance of existing works. In extracting process, by exploiting the SVM learning ability, the right secret-bits can be recovered. Based on each experiment, it shows StegSVM model has a better performance than the previous models. For further work, it is suggested that the other image domains through other intelligent methods should be investigated.