Data hiding in encryption–compression domain

This paper introduces a robust and secure data hiding scheme to transmit grayscale image in encryption-then-compression domain. First, host image is transformed using lifting wavelet transform, Hessenberg decomposition and redundant singular value decomposition. Then, we use appropriate scaling factor to invisibly embed the singular value of watermark data into the lower frequency sub-band of the host image. We also use suitable encryption-then-compression scheme to improve the security of the image. Additionally, de-noising convolutional neural network is performed at extracted mark data to enhance the robustness of the scheme. Experimental results verify the effectiveness of our scheme, including embedding capacity, robustness, invisibility, and security. Further, it is established that our scheme has a better ability to recover concealed mark than conventional ones at low cost.


Introduction
Nowadays, the growth of internet makes easier to disseminate multimedia information through various open channels [1]. Due to availability of internet, multimedia data can be easily tampered, stored, and shared through communication medium [2]. So, it leads to various problems such as copyright, unauthorized access, and security issues [3]. Watermarking is one of the highly recommended schemes to provide the protection of multimedia data [4,5]. The effectiveness of watermarking scheme can be evaluated using some important performance metric such as robustness, capacity and imperceptibility. Based upon on these performance metrics, watermarking approach can be divided into two parts such as robust and fragile [6]. In fragile-based watermarking scheme, it can be easily modified and it is used for content authentication and integrity verification [7]. On the other side, robust watermarking techniques are more robust against various attacks and it is normally used for copyright protection [8]. In general, watermarking scheme is classified based upon embedding domain of cover image [9].
We can categorize watermarking scheme into two parts such as spatial and transform domain. In spatial domain, watermark is embedded into cover by modifying the intensity of pixel image [10]. However, it is less resistant against various attacks. On the other side, mark is hided into transformed coefficient of cover image, which may greatly improve the robustness against attacks [11].
In this paper, LWT-HD-RSVD-based image data hiding scheme is proposed. Note that although various traditional image data hiding approaches have been reported, the interesting contributions of the proposed methods include the following four aspects: • We propose an image data hiding scheme based on LWT-HD-RSVD, which can provide both invisibility and robustness. LWT provides various advantages such as less distortion, low aliasing effect, very less memory requirement, low computation cost and reconstruction is very good [12]. The more precise components of the host image are obtained by HD [13]. This property of HD is used to provide high degree of robustness. It is reported that RSVD offered lower cost than SVD [14]. • The chaotic encryption [15]-then-wavelet-based compression scheme [16] is adopted to improve the security of the media data over possibly noisy network(s), while appropriate compression of encrypted data before transmission reduces the bandwidth demand.

3
• DnCNN is performed at extracted mark data to offer the additional robustness of the scheme. • The obtained results indicate that the method is satisfactorily invisibility, high payload and confirms its robustness against various attacks. Further, it is established that our scheme has a better ability to recover concealed mark than conventional ones at low cost.
Rest of the paper is organized as follows: "Related works" summarizes the related state-of-the-art techniques, followed by detailed description of the embedding, compression of encrypted data, and extraction procedure and de-noising of recovered data in "The proposed scheme". The results are discussed in "Experimental results" and the conclusions are summarized in "Conclusions".

Related works
Qingtang Su et al. [17] developed a copyright protection scheme for color images. Initially, Contourlet transform (CT) performed on each component of cover image and selected LL sub-band is decomposed into desired block. In their scheme, encrypted mark is produced then embedded in the transformed block to increase the security of the scheme. Their method was shown to be successful against against common attacks.
Chakraborty et al. [18] illustrated the comparative analysis of SVD-and RSVD-based watermarking approach in transform domain. First, this scheme decomposed carrier image into multiple sub-bands via DWT, and the selected sub-band is transformed using DCT. Further, RSVD is performed to modify the singular values of carrier image. The experimental analysis of this method proves that RSVD is much faster than general SVD-based watermarking approach.
Singh et al. [19] presented a DWT-DCT-SVD-based robust watermarking approach in transform domain. Initially, DWT is applied to decompose the host image into sub-bands. Further, DCT and SVD have been applied on selected sub-band. The watermark image is decomposed with the help of DCT and SVD. Watermark is inserted into transformed coefficient of host image. This scheme provides a better robustness against several attacks. Anand et al. [20] proposed a dual watermarking approach for smart healthcare system in compression-then-encryption (CTE) domain. First, cover image is transformed using redundant DWT and RSVD. In CTE scheme, it compresses the multimedia data before applying encryption technique to ensure security of multimedia data. Turbo code is applied to encode text watermark before the embedding process. In this procedure, compression and encryption of multimedia data are performed by wavelet-based compression and a stereo image encryption technique, which greatly improved the performance in several aspects. The author proposed a hybrid watermarking approach for digital images [21]. Initially, host image is decomposed into sub-bands using DWT. Further, selected sub-band of host image is transformed by DCT and SVD. The text watermark is encrypted before embedding process to enhance the security of this scheme. In embedding procedure, image and text watermark are embedded into different level of DWT coefficient of host image. Their method was shown to be successful against against common attacks. Authors have demonstrated a robust and secure watermarking approach for healthcare applications by Zear et al. [22]. First, Arnold scheme is adopted to scramble the mark image. Further, Hamming and Arithmetic encoding techniques are applied on signature and symptoms text watermark, respectively. In embedding process, encrypted image, compressed text and encoded text are embedded into different level of DWT coefficient of host image. Additionally, neural network is adopted to enhance robustness of the scheme. In [23], a dual watermarking scheme is used to enhance the security of digital contents. In embedding stage, it uses the second-level of DWT to decompose host image into different sub-bands. Further, selected sub-band is transformed using SVD. The encoded dual watermark is hidden into transformed coefficients of host image. The watermarked image is compressed via wavelet-based compression to reduce bandwidth demand.
A dual watermarking approach is developed for providing the security of medical application using DWT and SVD in [24]. Prior to embedding, Hamming code is adopted to encode the text mark, which may greatly reduce the channel distortion. The dual text and image watermark are embedded into transformed coefficients of host image. After embedding procedure, watermarked image is scrambled using chaotic encryption and then encrypted image is compressed via Huffman. This scheme provides the better results in terms of robustness, security, and imperceptibility. Author proposed an effective watermark approach for gray-scale image [25]. In this scheme, host image is transformed first into sub-bands using LWT and selected sub-band is transformed via SVD. The watermark image is also decomposed using fourth-level of LWT. In embedding process, watermark is inserted into transformed coefficient of LWT. After embedding procedure, digital signature is verified ownership authentication before watermark extraction procedure. This scheme provides better performance compared with some traditional watermarking scheme. Zheng et.al has implemented a robust watermarking method for copyright protection in transform domain [26]. Initially, cover image is transformed using DWT and DCT. Further SVD is performed to modify the singular value of transformed coefficient. The digital signature is applied in the embedding procedure to avoid falsepositive problem. This scheme provides the better robustness 1 3 against rotation attacks. In Ref. [27], author developed an efficient medical image watermarking in transform domain. Initially, DWT-SVD transformed host image and watermark is concealed into transformed coefficient of cover image. The chaotic encryption is performed on watermark to enhance the security of this scheme. A blind watermarking scheme is implemented to provide copyright protection of medical image [28]. In the first part of this scheme, DCT and Schur transform is performed to decompose host image and watermark is embedded into medium part of host image. In second part, DWT and Schur transformed used for embedding watermark into host image. So, this scheme provides better robustness and imperceptibility against various attacks.

The proposed scheme
The design proposed in this paper consists of four phases, i.e. (a) mark data embedding, (b) encryption and compression of marked data, (c) recovery of hidden data, and (d) de-noising of recovered mark. The main idea of the different sizes of mark embedding is to use LWT to decompose cover image through LWT-HD-RSVD. Then we use appropriate scaling factor to invisibly embed the singular value of mark data into the lower frequency (LL) sub-band of the cover. We also use chaotic encryption-then-SPIHT compression scheme to improve the security of the image over possibly noisy network(s), while the compression of encrypted data before transmission reduces the bandwidth demand. Additionally, DnCNN is performed at extracted mark data to enhance the robustness of the scheme. A simplified block scheme of different operations by the proposed solution is shown in Fig. 1. The detail description of mark data embedding, encryption and compression of marked data, recovery of hidden data, and the de-noising process of recovered mark is shown in the section "Embedding procedure", "Encryption and compression of marked data", "Extraction procedure", and "De-noising process of recovered data", respectively. The notations are summarized in Table 1.

Embedding procedure
In this process, cover image (C) and watermark image (W) are given as input to the embedding procedure. After embedding procedure, watermarked image C ′ is obtained as output. Algorithm 1 describes the embedding process in detail.

Encryption and compression of marked data
In this sub-section, watermarked image ( C ′ ) is encrypted using chaotic encryption to enhance the security of watermarking scheme. The encrypted image (Enc img ) is obtained by applying the XOR operation on chaotic key matrix and watermarked image ( C ′ ). Further, SPIHT compression is applied on encrypted image to reduce bandwidth and also save memory space. The SPIHT procedure contains three steps such as sorting, refinement and quantization for compress the image. Finally, compressed image ( Comp img ) is obtained as output. The detail steps of encryption and compression of marked data are explained in Algorithm 2.

Extraction procedure
Reverse embedding procedure is followed for extracting the watermark. Initially, Comp img is decompressed with the help of SPIHT decoding. After that, decrypted image Dec img is obtained by applying Chaotic Decryption on Decom img . In this process, decrypted image Dec img is given as input of extraction procedure and extracted watermark Ext wat is obtained as output. The extraction procedure of watermark is described in Fig. 1. The various steps of extraction process are described in Algorithm 3.

De-noising process of recovered data
In the proposed approach, DnCNN is implemented at extraction procedure to enhance the robustness and enhance visual quality of extracted watermark. Deep Learning toolbox is used to apply pre-trained denoising convolutional neural network. The several steps of De-noising process of recovered data are described in Algorithm 4.

Experimental results
All experiments done with the proposed scheme are simulated on a PC of 8 GB RAM using MATLAB R2019a. All used gray-scale host images with the size of 512 × 512 [29] are shown in Fig. 2. The mark images of varying size such as 256 × 256, 128 × 128 and 64 × 64 are shown in Fig. 3 [27]. We estimate the performance in terms of objective assessment is adopted in this paper, which is defined in Table 2.
The objective evaluation (PSNR, SSIM and NC) scores are depicted in Fig. 4. It can be seen from this figure, all the evaluation metric have high results. The validity of the proposed approach is verified for different cover images and variable size of watermark. The results obtained are summarized in Table 3.
According to Table 3, it provides performance for ten cover images and different size of watermark at gain value = 0.05. The highest PSNR value is obtained as 50.14 dB for Sailboat image at gain value = 0.05. The values of NC and SSIM are approaching 1 for all the cases. Diffusion key of logistic map   Peak signal to noise ratio (PSNR) [24] It determines the similarity between the cover and marked image PSNR = log 10 wherep(x, y) = luminance function, q(x, y) = contrast function and r(x, y) = structure function Normalized coefficient (NC) [30] It finds the similarity between original and recovered watermark The value of NC lies between 0 and 1 and unified averaged changed intensity (UACI) [31] These are used to obtain efficiency of image encryption algorithm against several attacks If the value of NPCR and UACI is more, then it provides more resistance against numerous attacks where Y 1 , Y 2 = encrypted and decrypted images, F = the largest pixel value, T = total number of pixels value in cipher-text images Further, best values of NPCR and UACI obtained are 0.9964 and 0.3916, respectively. Notably, values of SSIM, NC, NPCR, and UACI of our implemented scheme are higher than 0.9952, 0.9954, 0.9957 and 0.2686, respectively. It is observed that if decrease size of watermark, then our imperceptibility performance is increased and robustness value is decreased, respectively. The quality of extracted watermark is evaluated, when different types of attack are performed on watermarked images. The implemented scheme is simulated on different gain value. The experimental result is depicted in Table 4. In this table, we found the highest PSNR and SSIM value are 50.93 dB and 0.9999, respectively at gain value 0.008. However, the NC value of extracted watermark is 1 when gain value is more than 0.05. It is observed that if we increase the gain value, then PSNR and SSIM values are decreased; however, robustness improves.
According to Fig. 5, it can be observe that our proposed scheme is examined against various attacks with different size of watermark. The robustness is tested against JPEG compression with various quality factors. The quality factor is indicated as compression strength. If the quality factor is increased, then NC value is also increased. The NC values of speckle noise are greater than 0.9362 for three different sizes of watermark.
In median and average filter, NC values are greater than 0.9859 and 0.9725, respectively. In salt and peppers noise, NC values are more than 0.9274 for three different sizes of watermark. The robustness performance of our scheme against Gaussian noise is greater than 0.9089 for three different sizes of watermark. The NC values of sharpening and Poisson noise are more than 0.9980 and 0.9717, respectively. Our proposed watermarking technique is robust against all the attacks except Histogram Equalization attacks. Therefore, from the above analysis, it can be identified that implemented scheme achieves optimal trade-off among robustness and imperceptibility.
The robustness performance of our implemented scheme, when compared with some mentioned techniques [19,20,24,27] against attacks are illustrated in Table 5. It is remarked that the implemented scheme provides the better . It is clearly indicated from figure that performance of our scheme is found to better in term of all the attacks under consideration. Lastly, subjective evaluation [24] is also adopted to evaluate the image quality, which is defined in Table 6. It indicates that the smaller gain has proven to be more suitable quality of marked image.

Conclusions
This paper described a robust and secure data hiding algorithm that utilizes LWT-HD-RSVD for embedding of mark data. A main interesting point of the proposed solution is the mentioned chaotic encryption-then-wavelet based compression scheme which enhances the security  Fig. 6 Graphical results for comparison with [19] after attacks of the media data over possibly noisy network(s), while appropriate compression of encrypted data before transmission reduces the bandwidth demand. Further, DnCNN is performed at extracted mark data to offer the additional robustness of the scheme. Obtained results verified the effectiveness of our scheme. Furthermore, our scheme is more efficient at low cost when compared with similar existing methods.

Fig. 9
Graphical results for comparison with [27] after attacks