• No results found

3. Methodology

3.4. Experimental Architecture

In order to verify the correctness of the above hypothesis, some experiments were designed and implemented to investigate the influences between original pixels and randomized pixels. Two effective solutions are presented; both are subject to the modification of image pixel value. One is based on the modification of image chrominance, which is the Chrominance Modification Algorithm

(CMA). The value of RGB in each pixel is affected by changing the value of chrominance. This is in YUV,

where, the value of U and V will be changed, the value of Y is kept constant, rather than changing the colour intensity directly as luminance is more sensitive for human eyes [28].

The other solution is based on the conversion of the image frequency domain, which is called

Presentation Domain Transform (PDT). The compression of JPEG images is a lossy procedure, the

quantization phase is an irreversible process, even if reverse engineered, the initial value could never be revealed [28].

pg. 76

These experiments were divided into two stages. The reliability and applicability of both solutions need to be verified first, then the results need to be analysed and evaluated. The specific experimental procedure is described as below:

Required Tools

In order to enhance the transparency of the experiments as well as deeply understand the experimental procedure, two tools were developed in Java with JRE System Library “JaveSE-1.7”. One tool is based on data hiding and the other tool is based on data extraction.

Algorithm

The LSB algorithm was executed in the embedding process. Similar to the majority of online open- source software, the algorithm (LSB 1) can be implemented through replacing data of the last bit of each pixel value with each message bit. In addition, in the developed tool, the replacement bit(s) in each pixel is controllable, from LSB 1 to LSB 8, to implement multiple bits embedded to increase the data hiding capacity. This implementation also can be used as a reference tool to detect the quality threshold that represents the image distortion for the traditional LSB algorithm.

Measure Threshold

With the constant increase of data size, the maximum bits available within LSB 1 is sometimes not enough to hide all of the data. Generally, in order to digest a large volume of data, increasing the number of bits replaced is an easier way to implement data hiding, although the image distortion becomes appreciable with the increase in replaced bits.

Therefore, the threshold of the traditional LSB algorithm should be measured for understanding its limitation further. As described by Fyffe [64], the visual threshold was identified at 3 bits LSB (LSB 3) through experimentation with 25 men and 25 women ranging from 20 - 60 years old. The subject viewed multiple visual examples of images with and without embedded hidden data.

Notably, as mentioned above, luminance is more sensitive for human eyes. Hence, the threshold result may not be very exact if an answer is given using visual detection only. The visual comparison between the original cover image and the stego image in “StAndrews Bridge.png” and “Lena.png”, with both images containing 100 KB of embedded data from a TXT file using the LSB 3 algorithm is shown in Figures 39 and 40. The data was embedded from the upper left corner of the image, and down the vertical direction. Obviously, in Figure 39, the stego image has appreciable blur, from the left side, especially in the sky part of the image. However, under same conditions with the same algorithm, the visual difference is negligible in Lena.png.

pg. 77

Figure 39: The visual comparison between the original cover image and LSB 3 stego image, in StABridge.png

Figure 40: The visual comparison between the original cover image and LSB 3 stego image, in Lena.png

Therefore, due to the influence of pixel intensity, the identification of a threshold is unreliable if the detection statistics is only from human visibility. In the next section, in order to enhance the credibility of the results, the threshold will be identified by using corresponding measurement units.

Measurement Units and Tools

Two measurement units will be used to verify the quality of image in the next sections; there are the Peak Signal-to-Noise Ratio (PSNR) and the Mean Squared Error (MSE). As mentioned before, the quality of images and MSE are inversely proportional, the higher value of MSE indicates a dissimilarity

pg. 78

between the compared images. However, the value of PSNR indicates the quality of an image, with a higher value indicating a better quality.

Firstly, these measurement units will be used to identify the threshold of the LSB algorithm, through calculating the value between the original cover image and n bit (n [1, 8]) LSB stego image. The significant point of detection can be seen as the threshold.

Next, these units can also be used to evaluate the performance of two presented solutions, to verify which one is better.

In addition, an online histogram tool [59] is selected to measure the distribution of the image colour intensity.

Experimental Structure

Prior to implementing the experiment, the relevant preparatory works were performed:

 Use different n to execute n bit(s) LSB algorithm in several cover images and use three different size TXT files as the secret (embedded) file. The cover images were selected from different frequency domain and spatial domain images.

 Collect all of the generated stego images and determine the threshold of LSB algorithms through statistical measurement. A histogram result is recorded for analysing the distribution change for the different sizes of embedding data.

In order to verify the applicability of presented solutions, the entire experimental procedure is divided into three phases.

- Phase 1: Pick all of imperceptible stego images in which the replaced bit n is under the range of the algorithm threshold, and implement presented solutions to these stego images respectively.

- Phase 2: Verify the applicability and reliability of these solutions through comparing data extraction results between the original stego image and a processed stego image.

- Phase 3: Identify the efficiency of these solution through benchmarking (measurement units).

In this experimental procedure, in order to confirm the distortion level of stego images using

the LSB algorithm, the threshold should be researched first as the main risk often comes from

the target image which is under this threshold. Otherwise, the distorted image will be

identified malicious, and be blocked by defence system. Then, these images will be used to

pg. 79

examine the applicability of the solutions by observing the data extracted result after

implementing the relevant solutions. If the result is unreadable and there is not a noticeable

distortion in the image, it would prove the usability of these solutions and research hypothesis.

Finally, accordingly through benchmarking the different solutions, the most effective solution

will be identified.

Analysis

The evaluation of both solutions will be analysed through comparison with different parameters, such as histograms, PSNR and MSE. Finally, all of the results will be combined to verify the performance of each solution, and determine which one is the most effective solution for DLP systems to prevent the application of steganography.

pg. 80

Related documents