A Robust Algorithm for Digital Image Copyright Protection and Tampering Detection: Employing DWT, DCT, and Blowfish Techniques

The burgeoning multimedia industry has resulted in the proliferation of digital media files, readily accessible via the Internet. This availability, however, has also facilitated the unauthorized acquisition of these digital contents, leading to copyright infringement and substantial losses for the original content owners. The ease of content modification and rapid transfer across the Internet further exacerbate these issues. Consequently, data piracy and copyright protection have emerged as critical challenges in preserving ownership rights [1, 2].

Digital image watermarking serves as a solution to these challenges, providing a technique for tagging digital media to establish and assert copyright ownership. This process involves the insertion of copyright information into a digital image in a manner that does not compromise the image's quality. Beyond embedding copyright information, the ability to extract this information is crucial in affirming the identity of the digital image.

One of the prevailing watermarking techniques is frequency-domain watermarking, which encompasses various forms such as Discrete Wavelet Transform (DWT) [3-6], Fast/ Discrete Fourier Transform (FFT and DFT) [7, 8], Discrete Cosine Transform (DCT) [9, 10], Principle Component Analysis (PCA) [11-13], and Singular Value Decomposition (SVD) [9, 14-17]. Typically, these methods transform host images into the frequency domain, where the watermark is discreetly embedded.

Pallaw et al. [18] proposed a robust watermarking scheme utilizing the Slantlet transform (SLT), randomized-singular value decomposition (RSVD), and Firefly algorithm optimization technique. In their approach, the original image undergoes SLT for frequency domain transformation, followed by the division of the lower-frequency SLT sub-band coefficients into 3×3 blocks. Each block is then transformed into U, S, and V matrices using the RSVD algorithm, with the encrypted watermark bits embedded into the S matrix.

Abadi and Moallem [19] introduced a robust color image watermarking algorithm that merges the DWT and DCT. The DWT is applied to the green channel of the host images, and the watermark is subsequently distributed in the DCT coefficient, selected by a particular secret key.

Furthermore, Su et al. [20] proposed an adaptive blind watermarking method based on the slant algorithm, aimed at enhancing robustness against geometric attacks. Kumar and Singh [21] developed an adaptive color image watermarking scheme using DWT, alpha blending, and entropy concepts to balance imperceptibility and robustness. Additionally, Sharma et al. [22] presented a hybrid watermarking scheme combining singular value decomposition, discrete wavelet transformations, and the artificial bee colony for embedding a scrambled color watermark.

Salama et al. [23] suggested a triple-channel encrypted hybrid fusion technique to improve the robustness, imperceptibility, and security of medical images. Liu et al. [24] introduced image watermarking algorithms based on DWT, Hessenberg decomposition (HD), and SVD transforms. Their embedding process applied DWT to the host image to produce several sub-bands, followed by HD on LL, and SVD on the created H and the watermark logo. The singular values of the host image and the watermark were then added with a scaling factor, optimized using the Fruit fly optimization algorithm (FOA).

While most of the previously mentioned techniques primarily focus on enhancing robustness and imperceptibility, they do not address the detection of unauthorized alterations in original images. In response to this, Liu and Yuan [25] put forth a dual-tamper-detection approach that both detects tampering and facilitates self-recovery. The watermark in this method is composed of two check bits and one recovery bit. Initially, the m-bit Most Significant Bit (MSB) image is extracted from the original, from which the Parity Check Bit Labeled (PCBL) generates the first check bit. The m-bit MSB image is then partitioned into blocks, each subjected to the MD5 algorithm to produce the second check bit. The SPIHT (Set Partitioning in Hierarchical Trees) algorithm is employed on the original image to create the recovery bit. The resulting watermark is subsequently embedded into the m-bit MSB image of the original image.

Nagm and Elwan [26] propose a novel algorithm that assigns a unique code to each pixel of a medical image, thus safeguarding patient information against deliberate attacks. Sulaiman and Altaei [27] introduce an Extreme Learning Machine (ELM) classifier that is trained on extracted textual features to identify image tampering. Srivastava and Yadav [28] present a technique that leverages texture descriptors to differentiate between original and forged images. Initially, the image is converted into the YCbCr color space, followed by the application of a standard deviation (STD) filter to the Cb and Cr channels to extract vital object details. A support vector machine then classifies the images as either original or tampered.

Zhao et al. [29] introduce a tampering detection technique that employs maximum entropy criteria. The Itti model [30] is applied to the original image to generate a saliency map, which is utilized as the watermark. This watermark is then embedded into the image using the DCT-SVD method. At the receiver end, the saliency map of the watermarked image and the watermark are extracted and compared to detect any tampering.

Siddharth Bhalerao et al. [31] propose a reversible image watermarking scheme for tampering detection based on the Region of Interest (ROI). The image is partitioned into ROI and RONI (Region of Non-Interest) sections. Within the ROI, the prediction-error expansion technique is utilized to embed authentication data. This ROI is then compressed and embedded into the RONI region using the difference histogram expansion technique. It's evident that these tampering detection techniques primarily aim to ascertain whether an image has been tampered with, without considering copyright protection.

This study introduces a novel algorithm designed to both preserve ownership and detect any unauthorized tampering with images. The proposed algorithm integrates two distinct techniques: one aimed at safeguarding ownership, improving imperceptibility, and enhancing robustness, while the other focuses on detecting illicit tampering in original images. Recognizing that DWT demonstrates robustness against geometric and noise attacks, and DCT is resistant to JPEG attacks, the first technique capitalizes on the strengths of DWT and DCT. It does so by concealing the watermark within the green component of the image to bolster the method's security and resilience against various attacks. A two-level DWT is applied to the green component of the RGB image, and the resulting LL₂ of HH₁ is transformed into the DCT domain to insert the watermark within the middle coefficients of the DCT, guided by a secret key. The second technique encrypts the blue component of the image using the Blowfish algorithm [32], substituting the least significant bits of the encrypted blue component with those of the red component of the host image, thereby generating a new red component. The new red component, the watermarked green component, and the original blue component are then combined to produce the resultant watermarked image.

The primary contributions of this study are as follows:

We propose an image forgery detection and authentication scheme capable of concurrent ownership preservation and detection of unauthorized tampering and illicit alterations.

The preservation of ownership is achieved by embedding a watermark within the middle coefficients of the DCT of the LL₂ sub-band of the green component of the cover image. This approach offers several advantages:

• Embedding the watermark within the DCT coefficients of LL₂ enhances the robustness of the proposed method against various attacks.

• Concealing the watermark within the high-frequency coefficients of the green component has a negligible impact on the perceptual quality of the image.

• The use of a distinct secret key to select the coefficients, emphasizing the middle range of DCT, bolsters the security of the proposed algorithm.

Unauthorized tampering and illicit alterations are detected by using the Blowfish algorithm to create a Digital Watermark Prints Matrix (DWPs) based on integrity protection, encryption, and steganography, which can be concealed within the cover images. This method also offers several advantages:

• Tampering can be detected even if the forgery block size is as small as one pixel, as the DWP is distributed across every pixel of the watermarked images.

• The DWPs have a minimal effect on image quality.

• The created DWP exhibits dynamic properties, dependent on the structure of the cover images.

• Since the DWPs are generated from the host image, there is no need to send them to the receiver.

In the remaining sections of this research paper, the extraction and embedding methods will be explored in more detail in Section 2. Evaluation metrics, dataset information, and simulation results will be presented in Section 3. Finally, Section 4 will offer the conclusion of this research.

3.1 Dataset

Six standard color images are chosen from different databases to be used as host images for evaluating the proposed method performance, as shown in Figures 3(a)-(f). These images are Barbara, Jetplane, Lake, Lena, Mandrill, and Pepper, of size 512×512. The watermark image is selected as a binary image of size 20×50, as shown in Figure 3(g).

The images are selected for various properties. These properties are a mix of colors, details, textures, as in Barbara, Mandrill, and Pepper, and smoothness, as in Lena.

3.2 Evaluation metrics

• Peak Signal to Noise Ratio (PSNR)

PSNR is used to evaluate the effect of the watermark embedding process on the watermarked image. Higher PSNR values reflect a low degree of watermarked image distortion. The PSNR is calculated as [34]:

$P S N R=10 \log _{10} \frac{M A X^2}{M S E}$               (1)

where, MAX represents the maximum power of the image and equals 255 which is the maximum pixel intensity. Mean square Error (MSE) represents the noise power and is calculated as the cumulative squared error between the host and the watermarked image as follows:

MSE $=\frac{1}{M N} \sum_{i=1}^M \sum_{j=1}^N\left(X_{i, j}-Y_{i, j}\right)^2$                 (2)

• Normalized correlation coefficient NCC

NCC is used to measure the similarity between the original and the extracted watermark image. The NCC value ranges from 0 to 1. If the NCC value is close to 1, this means that the two images are very similar; otherwise, the two images are dissimilar. NCC formula between two images of size N×M is given by:

$\operatorname{NCC}(\mathrm{X}, \mathrm{Y})=\frac{\operatorname{Cov}(\mathrm{X}, \mathrm{Y})}{\sqrt{\operatorname{Var}(\mathrm{X})} \sqrt{\operatorname{Var}(\mathrm{Y})}}$                    (3)

where, Cov (X,Y) is the covariance between the image X and image Y. Var(X) and Var (Y) is the variance of image X and Image Y respectively. According to the Cov and Var definitions, the NCC formula can be written as:

$\mathrm{NCC}(\mathrm{X}, \mathrm{Y})=\frac{\frac{1}{M N} \sum_{i=1}^M \sum_{j=1}^N\left(X_{i, j}-\bar{X}\right)\left(Y_{i, j}-\bar{Y}\right)}{\sqrt{\frac{1}{M N} \sum_{i=1}^M \sum_{j=1}^N\left(X_{i, j}-\bar{X}\right)^2} \sqrt{\frac{1}{M N} \sum_{i=1}^M \sum_{j=1}^N\left(Y_{i, j}-\bar{Y}\right)^2}}$                         (4)

3.png

Figure 3. Test images (a) Barbara (b) Jetplane (c) Lake (d) Lena (e) Mandrill (f) Pepper (g) Watermark image

3.3 Simulation results

Different experiments are conducted to evaluate the proposed method's performance; some evaluate the proposed system's robustness, while others detect unauthorized alterations and unapproved tampering.

3.3.1 Robustness analysis

First, the proposed system is assessed corresponding to different gain factor values that were used to spread the watermark inside the DCT coefficients as mentioned in the embedding process step 5.

Table 1. The effect of the gain factor on the proposed system performance in terms of PSNR and NCC

Host Image	Watermarked Image		Watermark Image
	PSNR		NCC
	k=2	k=25	k=2	k=25
Barbara	55.7314	43.565	0.8369	1
Jetplane	55.684	43.5627	0.8492	1
Lake	55.6918	43.5713	0.6559	1
Lena	57.2813	43.6333	0.9587	1
Mandrill	57.4999	43.64	0.7665	1
Pepper	57.4331	43.7298	0.9263	1

Table 1 tabulated the values of the PSNR and NCC for the water-marked and watermark images, respectively, at gain factors 2 and 25. It can be observed from the table at k= 2 that the quality of the watermarked image is higher than its quality at k=25.

On the contrary, the quality of the watermark logo at k=25 is better than its quality at k=2; thus, the best k value that achieves the best performance is selected empirically between 2 and 25.

Figure 4 presents the effect of the gain factor on the watermarked image and the watermark logo, respectively. It can be noticed from the figure that the gain factor effect on the watermarked image quality is a little bit small, and the watermarked image is visually the same at low and high gain factors. However, the watermark logo quality is visually better at high gain factor values. On the other hand, it can be noticed that the DWPs inserted in the red component do not affect the watermarked image and the PSNR is higher than 40 db.

6666666666667777.png

Table 2. The effect of different attacks on the proposed algorithm performance in terms of NCC

Barbara	Jetplane	Lake	Lena	Mandrill	Pepper
Salt & Pepper Noise (Variance = 0.02)
0.9732.png	2.png	3.png	4.png	5.png	6.png
0.9732	0.9743	0.9598	0.9721	0.9754	0.9709
Salt & Pepper Noise (Variance = 0.005)
2-1.png	2-2.png	2-3.png	2-4.png	2-5.png	2-6.png
0.9978	0.9989	0.9944	0.9978	0.9989	0.9944
Gaussian Noise
3-1.png	3-2.png	3-3.png	3-4.png	3-5.png	3-6.png
0.9464	0.9341	0.9352	0.9419	0.9296	0.9363
Histogram Equalization
4-1.png	4-2.png	4-3.png	4-4.png	4-5.png	4-6.png
1	0.9978	0.9989	1	1	1
Sharpening
5-1.png	5-2.png	5-3.png	5-4.png	5-5.png	5-6.png
1	1	1	1	1	1
JPEG Q=90
6-1.png	6-2.png	6-3.png	6-4.png	6-5.png	6-6.png
0.9966	0.9978	0.9966	1	0.9888	0.9788

Table 3. The unapproved tampering detection and its effect on the proposed algorithm in terms of NCC

	Barbara	Jetplane	Lake	Lena	Mandrill	Pepper	Tampering Detection
Tampering  2x2	7-1.png	7-2.png	7-3.png	7-4.png	7-5.png	7-6.png	detected
	1	1	1	1	0.984	1
Tampering   4x4	8-1.png	8-2.png	8-3.png	8-4.png	8-5.png	8-6.png	detected
	1	0.9888	1	1	0.977	1
Tampering 16x16	9-1.png	9-2.png	9-3.png	9-4.png	9-5.png	9-6.png	detected
	1	1	1	1	0.977	1

Table 4. Comparison of the proposed system with the existing methods

This study was funded by Korea Agency for Technology and Standards in 2022, project numbers are K_G012002234001, K_G012002236201, and by the sabbatical support from Gachon University in 2023.