A Google Colab-Based Multi-Transform Watermarking Framework for Image Authentication

The combination of administrative offices, learning institutes, and medical services increasingly requires a reliable approach to verify the integrity of documents since they increasingly use digital documents and scanned files. These documents can easily be modified using today's editing software in a way that does not leave any traces of the changes [1]. Image authentication is a kind of active method; digital watermarking has been demonstrated as one of the most robust approaches. Digital watermarking is a technique in which imperceptible information is embedded in a host image and extracted later on to verify the integrity of image [2]. A lot of work has been carried out in frequency-domain watermarking methods, using discrete cosine transforms (DCT), discrete wavelet transforms (DWT), and singular value decomposition (SVD) techniques to achieve a reasonable security level and robustness to various signal processing attacks like compression, noise, and geometric distortions. This allows distributing watermarks over frequency coefficients [3, 4].

Several practical problems continue to face the watermarking process for document-like images despite great achievements in the technology. Document images contain less visual redundancy compared to natural photographs, so it can be easily degraded if a moderate embedding strength is applied [5]. Many existing watermarking techniques only apply one transform domain as the method for security, whether it is DCT, DWT, or SVD, it has limitation in achieving invisibility and robustness [6]. The hybrid approach using DWT and DCT or DCT, DWT and SVD in the transform domains achieves better performance of both requirements [7]. But most of the approaches are designed for photographs or medical images while ignoring the characteristics of educational and administrative documents [8].

Looking back at the recent literature, some frameworks for cloud-based watermarking for healthcare applications have been proposed [9]. However, they are hardly able to present a reproducible cloud-based implementation for supporting multiple watermark types and embedding modes in a single framework. Furthermore, many methods were proposed to authenticate the documents based on cryptographic hashing and blockchain authentication [10]; however, few explored the combination of a multi-transform watermarking method with the practical and no-installation cloud-based process for educational documents such as certificates, transcripts and administrative reports [11]. There is a lack of a united system that integrates multi-mode frequency-domain watermarking methods, multiple watermarking formats, systematic robustness testing, and visual detection of tampering in the cloud environment [12].

It is a great advantage for the proposed framework to be applied in image-based document authentication scenarios, such as those faced by educational institutions and public administrations. Academic certificates, transcripts, degrees, scanned administration reports, employment certificates, official forms and other documents commonly spread on electronic channels can all serve as examples. Once the watermark is invisible, the integrity of the document will be verified through the watermarking mechanism, so that any falsification, replication or unauthorized modification can be detected while the original appearance is kept.

Compared to other works, which typically focus on the investigation of a hybrid watermarking scheme, a novel and replicable watermarking framework including five different transform domain modes for embedding different types of watermarks is constructed in a cloud Google Colab space. Watermark generation, embedding, extraction, authentication, attack simulation, performance assessment and tamper localization are combined into an end-to-end cloud workflow.

3.1 System model

The system implementation employs a standard digital image authentication framework. When an authorized user wishes to watermark an image, they upload it to the embedding module, which subsequently injects a private watermark prior to storing the image and distributing it across various media. The watermarked image can be sent through non-secure channels, after which it is subject to multiple alterations and manipulations. The authentication phase of the system inspects an image, which may have been modified. The extraction algorithm then reconstructs the embedded watermark from this potentially corrupted image and compares it to a reference watermark that was previously saved. The decision-making process regarding the image's authenticity is achieved by comparing a correlation metric (which includes NCC) to a pre-defined threshold: high similarity scores mean an authentic image, while low similarity scores indicate a tampered one.

3.2 Watermark generation

Two types of digital watermarks are supported in this system: fixed watermarks, which carry institutional logos and short textual information of the institution in order to establish a strong identity during watermark embedding; and dynamic watermarks, which carry QR codes and cryptographic hash values (SHA-256) representing a unique digital fingerprint of the file. Before being applied to the image in the system, the watermark is transformed into a matrix, converted to binary or grayscale, resized according to the size required to overlay the host image at the selected position in the transform domain, and finally embedded.

3.3 Multi-mode embedding and extraction

There are five possible watermarking methods used in the system. They all employ different methods to embed and extract watermarks from the image in a transform domain. All modes were labeled with the same names during implementation and experimental write-ups. The modes are defined as M = {DCT, DWT, SVD, DCTDWT, DCTDWTSVD}. The DCT mode implements a standard DCT over 8x8 blocks of the luminance component, where some mid-frequency coefficients of each block are modified to carry the watermark bits. The DWT mode utilizes a single-level DWT on the luminance component and applies watermark embedding on the chosen sub-band of the wavelet-decomposed image. SVD is used to decompose a chosen region or the luminance component of the host image into its singular values, which contain most of the energy and are highly stable; then the singular values are manipulated to hide the watermark information in SVD mode. For DCTDWT mode, the DWT decomposition is performed on the host image, then an arbitrary sub-band is selected and DCT is used for watermark embedding in the selected sub-band block. Finally, DCTDWTSVD mode combines DWT decomposition on the host image, DCT modification on chosen sub-bands and SVD on the selected component to embed the watermark across multiple transform domains such that high degree of imperceptibility is maintained, also making it the most robust method among the chosen ones. Algorithm 1 describes the watermark embedding process, whereas Algorithm 2 illustrates the watermark extraction and verification process.

The parameter embedding strength was empirically chosen from a set of preliminary experiments to balance imperceptibility with robustness. Values for = 0.03 were chosen for DCT, DWT, DCTDWT, and DCTDWT_SVD modes, while = 0.01 was chosen for SVD mode, as larger perturbations of singular values would result in clearly discernible degradation in the image. For DWT-based embedding, the HL sub-band was chosen, which provides a reasonable trade-off between robustness and visual fidelity. In the DCT-based modes, the watermark bits were embedded within the mid-frequency coefficients within each 88 block such that it does not cause perceivable degradation to the image but still offer resistance against compression attacks. For authentication purposes, an image is said to be genuine when NCC is 0.95; otherwise, the image is said to be tampered with. This threshold was obtained by conducting various experiments on the robustness of the system under the above-discussed attacks.

3.4 Attack simulation and evaluation metrics

The system generates a fast-detection scheme that verifies the extracted watermark by directly comparing it with the original one (W) using NCC and BER. The same module lets the user input typical digital attacks such as JPEG compression, additive Gaussian noise, salt-and-pepper noise, blurring, rotation, cropping and scaling. The watermarked image will be treated for every attack by watermark re-extraction and, recalculation of NCC and BER to determine the influence of the attack. The system generates a simple tamper-localization map by drawing red dots on locations suffering from worst watermark degradation showing some points of alteration if there exists.

To maintain equal experimental conditions during robustness evaluation, the parameters of attacks were kept constant. JPEG compression was carried out with a quality factor of 50, Gaussian noise with zero mean and variance 0.01, salt and pepper noise with density of 0.02, central cropping with area of 10% of image area, Gaussian blur with kernel size of 5 × 5, scaling factor of 0.8 and image rotation by 5 degrees. These attacks are commonly encountered during document dissemination.

Five metrics have been employed in the performance evaluation. Visual quality imperceptibility has been assessed with MSE, PSNR and SSIM. These metrics achieve good results if MSE is low, while PSNR and SSIM should be high (close to 1). NCC and BER have been used to determine its authentication and robustness properties. Successful watermark recovery appears through high NCC and low BER, whereas tampered results are detected by low NCC or high BER. Authentic class has been given to all the images whose NCC is equal or greater than the given threshold, while the rest of the images have been classified as Tampered. The value for each of the metrics calculated in this experiment has been compiled in the Results and Discussion.

3.5 System workflow

The full process of the suggested system is illustrated in Figure 1. The user will create the necessary project folders and a storage area before uploading a host image. After uploading a host image, the user is required to select one of four kinds of watermarks namely text and logo, QR code, hash-based pattern, etc. And one of five embedding methods, namely DCT, DWT, SVD, DCT_DWT, COMBO, etc. After the watermark embedding, by the quick extract function, the system will calculate NCC to test proper watermarking. Apply digital attacks, including JPEG, Gaussian noise, salt-and-pepper, cropping, rotating, and blurring, etc on the image, and then re-extract the watermark to recalculate NCC and BER. The system automatically saves the images, tamper maps, and metrics in the summary file. A final decision is made as authentic when NCC is greater than or equal to, and the others are tampered with.

1.png

Figure 1. System-level workflow of the proposed watermarking system

5.1 Experimental setup

The proposed system has been tested on 512 × 512 benchmark images, which have been stored as PNG images. We have used four types of watermarks: logo, text, QR code, and hash-based watermark. For every image and watermarks five embedding modes (DCT, DWT, SVD, DCT_DWT, COMBO) are utilized and the watermarked images were subjected to standard attacks (JPEG compression, Gaussian and salt & pepper noises, cropping and rotation). The imperceptibility is measured by MSE and SSIM, and NCC and BER used to measure robustness and authentication.

Two exemplary types of documents, such as images, were chosen as test data. Quality images represent digitally generated images with little or no compression artifacts, whereas quality images are produced by moderate JPEG compression and some scanner-related degradation, which try to imitate acquisition in a real administration system. The size of every image was normalized to 512 × 512 before embedding into the images for fairness.

5.2 Imperceptibility results

In this step of invisibility testing, we examined imperceptibility by considering two images. We analyzed five embedding methods with the same watermark and the same embedding parameters. Before processing the watermark images, they were all standardized with a common resolution. The data in this case demonstrated that MSE values for DCT DWT and DCT_DWT were extremely low, which indicated that the watermark image did not contain any visually discernible distortion and that for SVD and COMBO images higher distortions existed because of their stronger embedding methods. The value of the image quality did not influence the outcome because the algorithm with the preprocessing methods provided an equally effective performance.

Figure 3 illustrates the comparison of the two images and gives a perspective of minor differences between high-quality and low-quality watermark images of different embedding methods. Table 1 gives quantitative imperceptibility values for two images.

Figure 3. Mean Squared Error (MSE) comparison for high- and low-quality images across embedding methods

Table 1. Baseline imperceptibility for high- and low-quality images

5.3 Robustness of discrete cosine transforms, discrete wavelet transforms, and COMBO under common attacks

As a further measure of robustness, we examined the two embedding techniques (DCT_DWT and COMBO) with three common attack operations (Gaussian noise, salt-and-pepper noise, and JPEG compression) and with the high- and low-quality versions of each image. For each setup, the NCC, BER, and SSIM were calculated between the attacked and original watermarks. This allows for direct comparison of the two methods as well as showing the effect of image quality on the recovered watermark.

The behavior of the different attacks and image quality is identical as reported in Table 2 containing numerical robustness results.

Table 2. Robustness metrics (NCC, BER, SSIM) for DCT_DWT, and COMBO under common attacks

In Figure 4, it can be observed that NCC value for all attacks, for both image qualities, remains more than 0.97, BER is almost zero and SSIM is almost one. The observation of the results also proved that the DCT_DWT and COMBO are robust, and the developed system is lightly influenced by initial image quality. The discrepancy between high- and low-quality images is small, showing robustness with regard to the input image quality.

4.png

Figure 4. NCC comparison for DCT_DWT and COMBO across attacks and image qualities

5.4 Computational cost

The average time taken to embed and extract per image for each of the five embedding modes is shown in Table 3. Each scheme completes both embedding and extraction in less than one second on the test platform. The single-transform methods are faster, and the hybrid methods are slightly more expensive. The increase in time associated with DCT_DWT and COMBO is quite low and should be of no concern for many applications.

Table 3. Average embedding and extraction times per image for the five embedding modes

5.5 Comparative performance of embedding modes

This suggests that DCT_DWT shows a good balance of robustness and quality, but that COMBO offers high robustness but is more computationally intensive.

Figure 4 displays the complete performance results among five embedding modes for imperceptibility (PSNR, SSIM), robustness (NCC, BER), and complexity (embedding time). As far as the image imperceptibility is concerned, the DCT and DWT modes yield the highest PSNRs (close to 48 dB) whereas the SVD and COMBO modes result in a bit lower PSNRs but with higher SSIM. Under standard conditions, perfect recovery for watermark for all methods has been reached (NCC = 1.0, BER = 0). The hybrid DCT_DWT mode results in the optimum performance for visual quality and complexity whereas the COMBO mode increases the robustness but requires more execution time.

5.6 Tamper localization results

Besides numerical robustness measurements, system also offers visual tamper-localization output to indicate affected areas by noise/tampering. With a Gaussian noise added to the extracted watermark as shown in Figure 5, tamper map shows red dots distributed around the affected regions and a normal region is clear. This shows that not only can the system detect tampering, but also intuitively present affected regions.

5.png

Figure 5. Tamper-localization output showing red markers over the affected regions

5.7 Overall discussion

The system described proves to achieve the optimal tradeoff between imperceptibility and robustness. All methods in the baseline experiment produce imperceptible embeddings with tiny MSE, perfect watermark retrieval (NCC = 1, BER = 0), and a discrepancy between high- and low-quality images is trivial. In normal attack conditions, the DCT_DWT and COMBO exhibit comparable, steady NCC values and virtually zero BERs, which means that the two algorithms maintain excellent robustness and image quality variation still has an insignificant effect.

The tamper-localization output maps the affected areas in a red-dot map, offering a clear view of the attacked regions. All in all, it offers robust authentication and reliable localization. The experimental outcome suggests that the designed framework appears promising in the administrative and educational document authentication field owing to the combined effect of imperceptibility, robustness and reproducibility. Nevertheless, more evaluation based on larger benchmark datasets and various attack cases must be carried out before actual implementation is warranted.

The analysis of Table 4 shows the comparison of the three modes in question. Overall, DCT_DWT achieves an excellent compromise among the measures of imperceptibility, robustness and speed, while COMBO has the highest robustness but a small sacrifice in processing speed.

Table 4. Comparative performance of embedding modes

The DCT and DWT modes generate the best quality because the watermark energy is injected into a certain number of transform coefficients with only limited modification. The SVD mode results in slightly larger distortion due to singular value perturbation, yet benefits from improved numerical stability. The hybrid DCTDWT mode balances between imperceptibility and robustness by making use of both frequency localization and energy compaction. The COMBO mode offers further improvement in robustness using multiple-stage embedding, but at the cost of somewhat larger distortion. Therefore, the DCTDWT mode is considered as the best choice in terms of overall performance and it is optimal if robustness against attacks is more essential than visual fidelity, in which the COMBO mode outstands other methods.