Enhanced Security and Robustness in Color Image Watermarking Using DWT-HD and Quaternion SVD under Hybrid Attacks

There are many image security algorithms that have been proposed in the past to fulfill the requirements. Early days of image security methods like cryptography and steganography may lead to complexity and poor visual quality of the recovered output images. Watermarking-based methods have recently gained significant popularity and have proven secure. To improve resilience, DWT-SVD-based watermarking methods are frequently used in the literature. The fundamental advantage of DWT is that it maintains invisibility, while SVD makes it difficult to erase watermarks without harming the image. Several safe image watermarking techniques were suggested in the previous studies [35-38]. Bose and Maity [35] have used singular values for selected wavelet sub-band coefficients of image data, so this research proposes an image embedding procedure for a sparse watermark. Using the Dictionary Learning (DL) technique, the watermark image is rendered sparse. But DL-based methods are complex and vulnerable to attacks. Next, a watermark decoder designed around the concept of compressed sensing (CS) is created within the context of an alternating direction method of multiplication (ADMM). Yasmeen and Uddin [36] offered multi-level operations of DWT with SVD hybrid combinations that are utilized to imbed and extract watermark features. But performance is required to be tested under hybrid attacks.

Sivananthamaitrey and Kumar [37] have proposed a watermarking method with two watermarks implanted in a color image. It is reliable and has a large embedding capacity. Using the SWT in combination with SVD decomposition, a grayscale image is used as one of the watermarks to show who owns the host image (SVD). The LSB approach is used to include a binary watermark to identify tampering, but it has low robustness. Naffouti et al. [38] have also recently used the DWT and SVD values to provide an advanced image watermarking technique in this study. The study first subjects a grayscale cover image to DWT decompositions at the sender side, and then subjects the original HH (high-high) components to Eigen decomposition. A similar procedure is carried out on a watermark image in grayscale. Next, the inverse DWT (IDWT) is used to combine two unitary and one diagonal matrix to create a digital watermarked image. They stated that DWT-SVD is efficient for invisible watermarking.

Singh and Singh [39], in their article, propose a new reliable and blind watermarking system for copyright protection based on DWT-SVD and DCT with Arnold Cat Map (ACM) encryption. Unauthorized reading and false-positive detection watermarking security issues in SVD-based methods are fixed by the suggested solution, but using the ACM map is predictable. The middle singular value of each block with a size of 4×4 of the (DWT) sub-bands of the host picture contains the DCT transform coefficients of these MSB and LSB values. However, the DWT-SVD-DCT method is computationally expensive, and errors or noise are introduced during transformations. Zear et al. [40] have studied an algorithm for multiple watermarking for healthcare applications based on DWT, DCT, and SVD. For identity authentication, the suggested solution employs three watermarks—a color Lump image watermark, a doctor's signature or identification code, and patient diagnostic data as text watermarks. Luo et al. [41] have studied DWT along with HD combined with SVD decomposition as the foundations of the unique picture watermarking technique suggested in this research. The host image is first divided into a number of sub-bands during the embedding process using multilevel DWT, and the resulting coefficients are then utilized as the input for HD. The SVD is simultaneously used to operate the watermark. The method seems robust due to the hybrid combination, but it is simple and requires more effort.

Rajput et al. [42] have also proposed a scheme with HD and SVD with discrete wavelet modification that is used to breakdown the cover image and watermark image (DWT). Following the use of SVD, an optimal scaling factor is used to directly include each unique value of the watermark into the CI singular components. Nazir et al. [43] have researched to offer an improved and reliable watermarking technique paired with a 4D hyper-chaotic system, and we examine its performance by extending and differentiating the previous work. In this work, they contribute by using an improved algorithm to watermark and protect color images. This algorithm combines a 4D hyper-chaotic system with HD, SVD, and DWT transformations for decomposition techniques (IEFOA). The HD_SVD method offers higher capacity and improved imperceptibility.

Mannepalli et al. [44] have designed a brand-new hybrid design combining block-chain-based encryption and invisible image watermarking. ED of the DWT coefficient is used to achieve the watermarking. The L-level DWT transform is used to decompose the color image and provide multi-resolution coefficients. To produce the edge coefficients, the HH wavelet band is subjected to ED. The watermark embedding uses edge coefficients and dilation differences to increase resilience. A block-chain-based hash technique designed for color photos is used to encrypt the watermark image. Yao et al. [45] have created a new way to manage secure digital watermarking that is based on block chains and multiple bits. They suggest using a self-learning watermark embedding algorithm to add pictures and binary messages to a hidden space of pictures. The method offered resistance to a wide range of attacks (geometric transform, JPEG, noise, etc.).

Rathord and Rai [46] have proposed an improved hybrid transformation method that combines the DWT-SVD technique with the HD combination. The entropy maximization stage of pre-processing adapts the LAB color space. The entropy analysis is the basis for the color space adaptation. We use parametric analysis of NC, PSNR, and SSIM to check how well the suggested and existing methods work on color images. The effectiveness of current watermarking techniques may be compromised by numerous assaults. Sahila and Thomas [47] have proposed prohibiting illegal access to multimedia data like audio, video, and photos; digital watermarking confers authority. An SVD-AES-based digital picture watermarking technique is the foundation of our research effort, which enables secure digital image transfer over networks. SVD-based watermarking is used initially. An image is encrypted using the advanced encryption standard (AES) technique after watermarking. AES decryption produces the watermarked image once the encrypted image has been sent via networks. The AES-128 encryption algorithm is used in the suggested procedure. The algorithm exhibits resistance to copy-and-paste, copy-and-paste, and cryptographic attacks such as brute-force attacks.

Pulgam and Shinde [48] have used encrypted watermarks for medical imaging. Gafsi et al. [49] have provided an enhanced chaos-based cryptosystem that can quickly encrypt and decrypt confidential color images. A complicated chaos-based pseudo-random noise PRNG-based security is suggested. High-dimensional characteristics have presented new vulnerabilities unless appropriately planned. Zhang and Ding [50] have designed an enhanced chaos-based cryptographic system that could quickly encrypt and decrypt confidential color images. To provide a high-quality key with high unpredictability behavior, high entropy, and high complexity, a complicated chaos-based PRNG is suggested. The secret image is to be encrypted using an upgraded architecture that relies on permutation, substitution, and diffusion features.

Mokashi et al. [51] have studied the authentication of images that are robust against attacks. A hybrid watermarking approach made up of the DWT-DCT in combination with the SVD (DWT-DCT-SVD) is proposed. Here, DWT-DCT-SVD is used to obtain the singular values of watermarks 1 (fingerprint) and 2 (signature). To increase its dependability, durability, and originality, the image is enhanced with a combination of watermarking techniques and two biometric traits. A dual anti-counterfeiting technique for QR codes is provided by Xun et al. [52], employing a combination of encryption and digital watermarking schemes. First, an RSA-based encryption mechanism is used to encrypt the authentication data. The watermark itself is encrypted by RSA, but neither the embedding approach nor its embedding locations are traced. The overall summary of the secure image watermarking methods is presented in Table 2. It can be concluded that the use of the encryption standard in combination with watermarking may significantly improve the overall performance.

It is clear from the summary in Table 2 that the ASE is the most efficient encryption standard, but the speed concern is still a challenge as watermarking uses hybrid transformation, which makes the algorithm bulkier.

Table 2. Summary of the secure image watermarking methods based on the encryption standards

2.1 Related work to intelligent and hybrid attacks

Watermarking techniques must be resistant to a wide range of threats. In recent studies, the hybrid attack [53-56] has been considered for evaluation. Xiao et al. [53] have proposed an algorithm for watermarking applications using DWT based on FOA to eliminate the issue of transparency and robustness. FOA is simple and has proven efficient for optimal scaling. Awasthi and Srivastava [54] proposed a dual image watermarking technique for data security that utilizes lifting wavelet transformation (LWT), PSO-SVD, and JAYA optimization. The method seems computationally complex. And simply using SVD might require more robustness. Ahmadi et al. [55] have preset the dual-color image intelligent PSO-based optimization in the SVD domain watermarking. The PSO is stochastic and comparatively slow to converge. Laishram and Manglem Singh [56] have designed a blind watermarking approach for the medical images, considering the various hybrid attack analyses. ROI-based watermarking is used. Mehraj et al. [57] have designed watermarking for heritage multimedia and also considered hybrid attack evaluations for watermarking cases. It is found that watermark extraction and image reproduction are still ill-posed problems under hybrid attacks. This paper presents a new intelligent image watermarking scheme based on DWT and SVD using the human visual system (HVS) and particle swarm optimization (PSO). The cover image is transformed by one-level (DWT), and then the LL sub-band of the transformed image is chosen for embedding. To achieve the highest possible visual quality, the embedding regions are selected based on HVS. Ahmadi et al. [58] and Ahmadi et al. [59] present an intelligent DWT-SVD-based watermarking, including PSO optimization for invisibility improvement. The best HVS-quality regions are selected for embedding the watermark. A blind CIW approaches by watermarking for copyright protection. The watermark is proposed to be embedded in the B channel of RGB space. The method performs significantly well. Ahmadi et al. [60] have added the additional digital signature for optimally watermarking implementations. Watermark schemes have discovered and developed their advantages by means of singular-value decomposition. Therefore, in recent decades, there has been a significant increase in the use of the SVD domain in many new image watermarking schemes, which show high robustness and imperceptibility. The material in this article is crucial and has the potential to help researchers create effective, resilient, and hybrid SVD-based picture watermarking systems. Sharma et al. [61] have proposed a nature-inspired secure, and intelligent approach to image watermarking. They have intelligently scrambled the watermark data with the use of a chaotic map for embedding. Prabha and Sam [62] have adopted grasshopper optimization, which is in a chaotic domain for secure watermarking design, but the method looks computationally complex. Hosny and Darwish [63] have used fractional order moments using the chaotic map for watermark embedding. Hasan et al. [64] have used the DWT-DCT combination for securing watermark information. Kazemi et al. [65] have proposed an ML-based CIW approach using NN-based computation. Garg and Kishore [66] used the PSO to find the optimal watermarking solution. Recently, Yamni et al. [67] proposed the use of Artificial Bee Colony (ABC) optimization to improve the effectiveness of the CIW system, stating that optimization can be effective in selecting the optimal scaling factor.

In this paper, it is proposed to use a hybrid combination of the HD and SVD decompositions. The mathematical modeling of each of these methods is presented in this section.

HD: represents the Heisenberg Decomposition, it is a matrix decomposition used for only square matrix decomposition [13, 14].

A nxn square size matrix X may be decomposed using HD as given in Eq. (2).

$P H P^T=H D(X)$             (2)

SVD: The SVD method is a mathematical factorization algorithm used for real complex image matrixes. Using SVD, the cover image is decomposed into a wavelet sub-band matrix $C i$ of size $m^* n$, represented as the product of three matrices containing scalar values as follows:

$C_i=U_C * S_C * V_C^T$                (3)

where, S_C represets the singular values of the cover image. The values of S_C are all non-zeros and are placed diagonally in the matrix. The use of SVD for the watermark insertion process is robust and less affected by most of the attacks. Similarly, the inverse process is also the multiplication of all three decomposed matrices together. Therefore, SVD is most widely being used for watermarking digital images.

In this section, the expected outcomes of the proposed watermarking methods are presented. This paper has preset the results in two passes. In the first pass, the results of the proposed watermarking using dual-key DWT-HD-SVD-dual are presented and validated against the work of Luo et al. [41]. The results of the watermarking are also compared with those of Rathord and Rai [46]. In the second part of the section, the parametric evaluation of the performance is also presented. Additionally, proposed secure watermarking with encryption is implemented, and qualitative and quantitative results are preset. The attack analysis of the proposed encrypted watermark is carried out for the logo of the study [43], and the results are compared. The two input color images taken from reference papers are used for the current study’s presented in Figure 6. The two kinds of logo images are also shown in Figure 6 (c) and Figure 6 (d), respectively, from the researches [41, 43].

The first stage of basic watermarking and extraction results is shown in Figure 7 for the Lena image with the L component. The efficiency of the watermark retrieval can be clearly observed. The comparison of watermark extraction results for Rathord and Rai [46] with our proposed dual key-based wavelet watermarking is given in Figure 7. It can be observed from the figure that the proposed approach clearly outperforms the existing approach and recovers the closest approximate watermark. In addition, the watermarked image's visual quality is also improved. The attack analysis is carried out using the FOA. Using optimization may improve the invisibility and robustness criteria of watermarking. The performance of watermarking is evaluated under the exposure of attacks, viz., filtering, different types of noises, cropping attacks, in addition to JPEG compressions, sharpening, motion blur, and rotation.

6.png

Figure 6. Host and the watermark input images used for evaluation in the recent studies. The Host images of 512×512 and watermark of 356×356 sizes are considered

7.png

Figure 7. Comparison of the extracted watermark using the Rathord and Rai [46] and proposed dual key based Watermarking

6.1 Performance evaluation under attacks

The proposed method's performance is evaluated using 12 different types of attacks on the watermark images. To evaluate the performance, statistical parametric features are extracted and evaluated under various attacks.

The basic properties and the parameters used for the various attacks are shown in Table 3. The mask sizes and variances need to be observed. The mathematical values of the parameters are defined as follows:

Table 3. Description of the attacks used for comparative analysis [43]

8.png

Figure 8. Validation of results for watermarking with Luo et al. [41] for the Lena image under various attacks

To further validate the code's effectiveness, the watermark extraction is validated under various attacks, as shown in Figure 8. Validation is shown for watermarking with Luo et al. [41] DWT-SVD-HD for the Lena image. It can be observed that watermarks are successfully recovered in all cases of attacks.

Figure 8 compares Luo et al. [41]'s validations of the performance of the watermarking results. Figure 8 demonstrates that the proposed dual-key-based secure watermarking method is capable of efficiently recovering the watermark under various noise and filtering attacks, including sharpening and histogram equalization. Only slight performance degrades in the case of large motion blur concerns.

A. Parametric performance under attacks using FOA

To demonstrate performance under various attacks, this section presents the variation of PSNR performance across different scaling parameters, which are plotted against different attacks. The FOA optimization, as proposed in Algorithm 4, is employed to determine the optimal scaling parameters α\alphaα for improving robustness during the attack analysis.

B. Quantitative PSNR performance comparison

The performance comparison of the achievable PSNR values for the Lena image is shown in Table 4 for the three methodologies. Table 4 shows that the proposed method performed well in PSNR compared to the existing methodology of Liu et al. [14], but due to the use of the dual key, more robustness is observed. The proposed method improves the SNR over the Liu et al. [14] by approximately 0.11 dB.

Table 4. Performance comparison of PSNR for existing and proposed method based on DWT-HD-SVD for Lena image

9.png

Figure 9. The PSNR improvement in dB with DWT-HD-SVD dual key method

10.png

Figure 10. Results of the PSNR comparison for different attacks with proposed method

11a.png

11b.png

Figure 11. Results of the NC calculated by the proposed dual key based method under the different attacks

$\operatorname{Psnr}=\frac{\sum f(x, y)^2}{\sum(f(x, y)-W(x, y))^2}$             (17)

The extracted PSNR offers a significant improvement over the dual-key method of around 0.6 dB. The parametric performances of PSNR improvement in dB are shown in Figure 9 and are evaluated and plotted in bar chart form for PSNR.

Figure 10 shows that the proposed dual-key-based security method has a better capacity for watermark extraction. It is observed that the proposed method outperforms the existing one for different watermark sizes (128×128 and 64×64). Figure 10 displays the PSNR comparison results for various attacks using the proposed method. In this paper, Gaussian filter, median filters, Gaussian noise, average filter, and motion blur attacks are considered for evaluating the performance under the presence of Gaussian filter. Attacks for different attacks are presented in Figure 10.

Figure 11 shows that for the motion blur attack, the NC value is the lowest, as shown by the yellow color mark. But overall, all NC values terminate at more than 0.95, which is good performance under attacks. Figure 11 also compares the NC values of our proposed method with those of Nazir et al.'s NC study [43]. It can be clearly observed that the proposed method outperforms all attacks and offers higher NC values. It can be observed from the red circle that the NC variation is significantly less for all attacks in our proposed method, which justifies the performance improvement over the existing approach.

$N C(x, y)=\int f(x, y) * f(x-t, y-t)^{\prime}$              (18)

Similarly, the results of the SSIM of the proposed method are validated and calculated for the proposed dual-key-based method for different attacks.

$\operatorname{SSIM}(x, y)=\frac{\left(2 \mu_x \mu_x+C_1\right) *\left(2 \sigma_{x y}+C_2\right)}{\left(\mu_x{ }^2+\mu_y{ }^2+C_1\right)\left(\sigma_x{ }^2+\sigma_y{ }^2+C_2\right)}$               (19)

It can be observed that the proposed method offers significantly higher SSIM values, as using dual-level security to add more robustness and a and a slight compromise to SSIM values is considerable. The Gaussian noise performance is at its minimum for SSIM, and it is even in the range of above 0.7. The performance improvement with the proposed LAB-Dual Key-based watermarking can be clearly observed from the comparison of performance under the different attacks. For proposed and existing methods, as shown in Figure 12, there is a significant improvement in the SSIM performance.

12a.png

Figure 12. Results of the SSIM comparison for different attacks with proposed DUAL key based DWT-SVD-HD dual method

6.2 Secure encrypted watermarking outcomes

This paper has proposed a rapid encryption standard for securing the watermarked image during transmission or storage. This section presents the qualitative and quantitative evaluation of the secure watermarking results of the proposed method. While in the second pass of the paper, the attacks are applied to the encrypted images and reconstruction is carried out, and the results are compared with the method of the paper [43]. An example of the various attacks applied to the Lena image is shown in Figure 13.

13.png

Figure 13. Results of our Dual key embedding under JPEG attack with FOA optimization

It can be clearly observed from Figure 13 that the NC values are evaluated under the variation of different noise variances and the compression ratio, quality factor, and window sizes. The proposed method offers significant NC values under various noises and even for different image sizes. Three image sizes (256×256, 128×128, and 64×64) are considered for the evaluation in the paper.

It is observed that NC is maximum for the 256×256 size, and as the quality factor might increase, the overall NC value is achieved to be around 100 percent. The performance under JPEG compression 2000 is evaluated for different compression ratios. It is found that at the lower compression ratio, the NC value is near 1, and as the CR increases, the minimum NC is offered as 0.995 for the 64x64 size, which shows the effectiveness of the proposed DWT-SVD-HD-Dual method. The low-pass filter performance of NC improves as the size of the image decreases.

6.2.1 Qualitative comparison with the previous study [43]

The extracted results for crypto weights for the encrypted image under different noise attacks are shown in Figure 14. It can be observed that, with significant invisibility and shuffling, the encrypted data is robust and secure. An example of the sequential outcomes is shown in Figure 14, plotted using the proposed method and the watermark by Nazir et al. [43] for salt and pepper noise attacks.

The quality of the encryption remains consistently high even under various noisy attacks, as observed from the crypto weights in Figure 15. The proposed method performs equally well for different kinds of attacks in terms of crypto weights. Since the attacks are applied in the encrypted domain, the quality of the histogram equalization attack is superior in this method.

14a.png

14b.png

14c.png

Figure 14. Sequential results under secure watermark for Salt & Pepper Noise attacks

15a.png

Figure 15. Results of proposed DWT-SVD-HD watermarking under various attacks

16a.png

16b.png

Figure 16. Results of the extracted watermark under the AES secure encrypted watermarked image exposed to various attacks

6.2.2 Attacks analysis for watermark with Nazir et al. [43]

The results of the watermark attack analysis applied to the encrypted images are evaluated in Figure 16. The results are preset for the extracted watermark under the AES secure encrypted watermarked image exposed to various attacks. The quality of the extracted watermark is clearly observed, and even in the worst case of motion blur, the watermark is visible in Figure 16 (e), justifying the efficiency even considering the higher level of noise in the worst cases.

6.3 Evaluation of hybrid attacks

Another experiment was carried out to justify the performance under the evaluation of hybrid attacks.

There are two independent cases of hybrid attack analysis considered in this paper: 1) simple watermarking + attacks without encryption; 2) watermarking + AES encryption + attacks.

Case 1: Hybrid attacks without encryption.

The results of the evaluation of secure watermark extraction under the exposure of various dual hybrid watermark attacks are presented in Figure 17.

In this case, hybrid attacks are applied to the watermarked image. Figure 17 shows the evaluation of results for watermark extraction. Using double hybrid attacks as combinations shown in Figure 17 (b-e), The low-pass filter and the noisy attacks are considered in the case of the rotation attack for analysis.

In any case, only two attacks are applied simultaneously. It can be concluded from the figure that for the PSNR of 27.91420 without attacks, the respective PSNR for each attack case is shown in Figure 17.

For noisy attacks, a significant PSNR in the range of 25 to 26 is achieved even in hybrid attacks. But in the case of a rotation attack, PSNR is reduced to 18 dB. But the quality of the recovered image is relatively visually pleasing.

Additionally, multiple hybrid attacks are applied to the watermarked image. Figure 18 shows the evaluation of the results for this case. Four attacks, including rotation, Gaussian low pass filter, noise, and JPEG compression, are applied in the combinations shown in Figure 18.

The types of noisy attacks are changed from Gaussian, salt & pepper, and speckle noise, keeping other attacks the same as rotation and compression attacks. Results corresponding to the different noise cases are shown in Figure 18.

It can be concluded that in multi-hybrid attacks, PSNR is reduced from 27.91420 to 18 dB, and the quality of reproduction is still justified by the NC values as shown in Figure 18.

17.png

Figure 17. The result evaluation of secure watermark extraction under the exposure of the various dual hybrid watermark attacks

18.jpg

Figure 18. The result evaluation under the various multiple hybrid attacks including JPEG, Rotation, Gaussian filter, and Noise. a) Original Image, b) watermark extracted

19.png

Figure 19. The result evaluation under multiple hybrid attacks applied on Encryption domain

Table 5. Comparison of the PSNR for the watermark extraction under hybrid attacks

Table 6. Comparison of the PSNR and NC for watermark extraction under various hybrid attacks using compression

Case 2: Hybrid attacks applied to an encrypted image

As a special case, hybrid attacks are applied to the encrypted image instead of the watermarked image. Figure 19 shows that, in the case of dual hybrid attacks, the performance is relatively stable.

$H_{a t k}=A 1+A 2+\cdots . . A n$                   (20)

$A_1=G L P_{a t k}, A_2=G N_{a t k}, A_3=R o_{a t k}, A_4=\operatorname{Comp}_{a t k}$                 (21)

It can be observed from Figure 19 that although watermark extraction is still possible under hybrid attacks, for several hybrid attacks, including filtering, noise rotation, and JPEG compression, the performance of image recovery is still an ill-posed problem.

The PSNR performance of reconstruction is shown under hybrid attacks in Table 5. It can be observed that for noisy and filter attacks, the performance is nearly identical to the true image case, with less than 4% error in PSNR. This reflects the efficiency of the proposed method in the noisy case.

The qualitative comparisons of parameters are given for the two cases of hybrid attacks as shown in Table 6. It can be observed that using the encrypted watermark may offer a higher recovery PSNR for all attack cases. Thus, the proposed secure watermarking is more robust under noisy attacks using AES encryption.