Blind Dual Watermarking Scheme Using Stucki Kernel and SPIHT for Image Self-Recovery

In this paper we propose a blind dual watermarking scheme using Set Partitioning in Hierarchical Trees (SPIHT) and Stucki Kernel halftone technique for the tamper detection and image self-recovery. The watermark consists of authentication bits for tampering area location and recovery bits for image restoration. We generate two recovery bits to ensure the high-quality recovery of the tampered image. The primary recovery bit is generated by the SPIHT encoding, and the secondary recovery bit is generated by the Stucki Kernel halftone technique. Then the authentication bit is generated based on the recovery bits. Before embedding the watermark, we shuffle the watermark bits through Arnold cat mapping and diagonal mapping to improve the security and quality of the restored image. LSB-based watermarking technique is used to embed the watermark into the original image to ensure the invisibility of the watermarked image. Experiments have been conducted on two datasets, BOW2 and USC-SIPI, and results show that the proposed scheme can achieve high restoration quality. Comparison with the existing works demonstrate the good performance and superiority of the proposed scheme.

tampering digital images has become a serious problem. 23 The watermarking techniques can be classified into different 24 categories: fragile watermarking, which is designed to detect 25 every possible change in the watermarked image, and it is 26 very sensitive, thus suitable for verifying the integrity of the 27 data. Semi-fragile watermark, which enables the embedded 28 information to withstand acceptable operations such as com- 29 pression, slight noise, filtering, A/D or D/A conversion, thus 30 it is often used for robustness checks. Reversible watermark-31 ing, which is also known as lossless watermarking, fully 32 The associate editor coordinating the review of this manuscript and approving it for publication was S. K. Hafizul Islam . extracts the embedded information and fully restores the 33 original image. And robust watermarking, which is to ensure 34 the integrity and security of the watermark information under 35 various attacks. 36 In recent years, watermarking scheme for self-recovery 37 has become an important research topic. In the authentica-38 tion process, the watermark is extracted and compared with 39 the content information of the original image, to determine 40 whether the content of the image has been tampered. It can 41 not only judge tampered image and its tampered area, but also 42 restore the tampered area. The researches focus on ensuring 43 the invisibility of watermarked image and improving the 44 localization of tampered area and performance of image self-45 recovery. In these works, the watermark is generated from the 46 host image and then embedded. By extracting the watermark 47 information, the tampered region can be detected and located 48 by the authentication bits. On this basis, the detected region 49 can furtherly be recovered by recovered bits. Therefore, these 50 schemes can be classified as fragile watermarking. 51 The idea of restoring fragile image watermarking algo-52 rithm was first proposed in [1]. The purpose of the 53 watermark data into host image may cause some distortion 95 to the original image. But it also brings the capability of

105
Although it is resistant to attacks of JPEG compression 106 with the quality factor above 75, its PSNR is only 35dB. 107 Zhang et al. [12] proposed a tamper detection, localization, 108 and recovery scheme for encrypted images using DWT and 109 compressed sensing. If the watermark or low frequency parts 110 are tampered, the recovery accuracy is 100%. Bravo et al. [13] 111 introduced a receiver that used authentication bits to locate 112 changed blocks of pixels and then performed an iterative 113 recovery mechanism to compute the original values of the 114 watermarked pixels. In this method, it embeds reference 115 bits and some authentication bits to the 3 Least Significant 116 Bit (LSB) planes of the image. Therefore, based on the 117 premise of being less perceptible and resistant to attacks, 118 we choose to use the method of watermark embedding to 119 LSB. Haghighi et al. [14] presented a fragile blind quad 120 watermarking scheme, which was proposed for image tam-121 per detection and recovery based on wavelet transform and 122 genetic algorithm. This scheme introduced two techniques 123 called Mirror-aside and Partner-block, and it can achieve 124 the average PSNR and Structural Similarity Index (SSIM) 125 values of the watermarked image of about 46 dB and 1. 126 Su et al. [15] proposed an effective self embedding fragile 127 watermarking scheme for medical images. The self-recovery 128 information and verification code of each block were pre-129 generated, and then embedded into other blocks with the 130 help of tortoise shell and embedded table. The simulation 131 results showed that the accuracy of tamper detection can 132 reach 99.83%, and the average PSNR of the restored image 133 can reach 42.11 dB. Hamid et al. [16] introduced a method of 134 extracting local features in DCT domain, where the positions 135 of three DCT peaks were checked to distinguish 13 texture 136 contours with different number of edges, edge directions, and 137 combinations of the two. Sinhal et al. [17] proposed a blind 138 fragile watermarking scheme for color images which can 139 provide effective image tamper detection and self-recovery. 140 The pseudo-random binary sequence based on the key was 141 used as a watermark for tamper detection. Experimental 142 results showed that the scheme achieved nearly 99% accurate 143 tamper detection and significant tamper image restoration. 144 Shehab et al. [18] proposed a new image authentication 145 and self-recovery scheme for medical applications based on 146 fragile watermarking. Singular value decomposition (SVD) is 147 used to calculate the transformation in the original image by 148 inserting the trajectory of block level SVD into the LSB of 149 image pixels. We summarize the existing works in Table 1. 150 It can be seen that most of the algorithms have the prob-151 lems of low PSNR of recovered images and small attack 152 resistance area.

153
In order to further improve the accuracy of tamper detec-154 tion and ensure the quality of self-recovery, in this paper 155 we propose a blind dual watermarking scheme using Set 156 Partitioning in Hierarchical Trees (SPIHT) and Stucki Kernel 157 halftone technique for the tamper detection and image self-158 recovery. The SPIHT-based recovery bit is used to recover 159 the tampered area as an image digest, while the halftone-160 based recovery bit can provide a second chance of recovery 161 and additional information for authentication. At the same 162 time, we improve security by permuting the coefficients of 163 VOLUME 10, 2022 the recovery bits. In this proposed scheme, we generate two 164 recovery bits to ensure the high-quality recovery of the tam-165 pered image. The primary recovery bit is generated by the 166 SPIHT encoding, while the secondary recovery bit is gener-167 ated by the Stucki Kernel halftone technique [19], [20], [21]. 168 On this basis, we then generate the authentication bit by 169 applying the logical operation. Furthermore, we employ the 170 Arnold's Cat Map (ACM) [22], [23]

180
The goal of the scheme is to detect the tampered region There are two types of watermarks that make up this embed-190 ded part. The watermark bits for recovery, which are for 191 the purpose of recovering the tampered regions, and the 192 watermark bits for authentication, which are for the pur-193 pose of tampering localization, are created and embedded 194 into the host image. In this work, we propose to generate 195 the recovery watermark using the Stucki kernel and SPIHT, 196 on the basis that SPIHT can be used as image digests, thus 197 to recover the tampered regions, while the Stucki kernel can 198 provide a second chance for recovery and can also be used 199 as authentication information. Additionally, we propose to 200 apply ACM and Diagonal Mapping to shuffle and encrypt the 201 coefficients in each block, which in order to increase security. 202 The process of generating and embedding the watermark is 203 shown in Figure 1. After producing the recovery bits, we use 204 a logical operation between the two recovery bits to generate 205 the authentication bits, providing a detecting method with 206 high precision and credibility. Afterwards, we concatenate the 207 recovery bits and the authentication bits after shuffling them 208 by ACM. Finally, the watermark composed of recovery bits 209 and authentication bits is embedded into the LSB planes of 210 the host image.

211
To generate the recovery bits, the SPIHT and Stucki ker-212 nel are respectively applied to the host image respectively, 213 to generate the primary recovery bits W Rec1 and the secondary 214 primary bits W Rec2 . The SPIHT is widely used as embedded-215 compression algorithm in digital signals compression. SPHIT 216 can generate a wavelet transform coefficients bitstream at 217 our desired rate and can be used to reconstruct the image 218 even the received bit stream is interrupted anywhere with 219 good progressive transmission. Therefore, in our method, 220 we employ the SPIHT for generation of the primary recovery 221 bits W Rec1 .

222
The SPIHT sorts the rounded multi-resolution wavelet 223 transform coefficients according to their magnitudes and 224 transmits them based on significant bit order for higher 225 quality because the quality of reconstruction depends on the 226 output rate exploited [24], [25]. The same process will be used 227 availably to the decoder inversely as well. These similarities 228 can be found through wavelet transform spatial orientation 229 trees as shown in Figure 2. To satisfy our requirement, we set 230 the compressing rate as 1.5 bpp, and the length of bit stream 231 is 393172, given the host image be 512 × 512. As a result, 232 the SPIHT matrix of 256 × 256 will be generated after zeros-233 padding and resizing.

234
Then we propose to apply the mapping to the generated 235 matrix, to improve the security of the watermarked image 236 and furtherly quality of the recovered image. In this work, 237 we employ the ACM and diagonal mapping to shuffle the 238 result. The specific steps are: divide the matrix into four equal 239 columns from left to right, denoting as p1, p2, p3, and p4. 240 If the position of the coefficients that generated by ACM is 241 p1, then we shuffle it to p3, so as p2 and p4. And the same 242 operation will be applied to right side of the matrix. In this 243  With the generated primary recovery bits W Rec1 , and sec-261 ondary recovery bits W Rec2 , the authentication bits are then 262 calculated accordingly. Given the host image of size M × N , 263 firstly the blocking is applied to the host image to generate 264 blocks of b × b, from each of which one authenticate bit will 265 be calculated. In the following experiments, we use b = 2 in 266 the proposed scheme. That means, one bit will be generated 267 as authenticate bit in each 2 × 2 block. By representing the 268 generated primary recovery bits W Rec1 in its binary form, 269 VOLUME 10, 2022   the watermark information will be embedded into the two 285 LSB planes of the host image. Figure 4 shows the demon-286 stration of watermark embedding in one block of 2 × 2. 287 In order to resist attack of copy-move, we propose to apply 288 the ACM into the watermark information as well before being 289 embedded.

291
In order to detect and recove the tampered regions, the water-292 mark information is extracted from the received watermarked 293 image. The ACM will be used to generate the recovery 294 bits and authentication bits, same to how watermark gen-295 eration was mentioned above. Once the recovery bits have 296 been retrieved, the authentication bits may be computed, 297 much like with the watermark generation process. We can 298 identify the tampered regions by comparing the computed 299 authentication bits with the extracted authentication bits. 300 On basis of that, the extracted recovery bits can be employed 301 for the recovery of the tampered regions. Figure 5 shows 302 the procedures of detecting and recovering the tampered 303 regions.

304
Similar as described in Section II.A, firstly, the received 305 image is blocked into 2 × 2 blocks, from each of which, 306 the watermark information is extracted by extracting the 2 307 LSBs of each pixel. After permuting and decrypting the 308 extracted information by ACM, eight bits can be obtained 309 from each block. In each block, the first six bits are the 310 primary recovery bits, notated as W Ext_Rec1 , while the sev-311 enth bit is the secondary recovery bit, notated as W Ext_Rec2 , 312 and the eighth bit is the extracted authentication bit, notated 313 as W Ext_Auth . From the extracted recovery bits, the corre-314 sponding authentication bits W Cal_Auth can be calculated 315 using Eq. (1), (2), and (3). From W Ext_Auth and W Cal_Auth , 316 the tampered regions can be detected by (4). It is noted 317 that the morphology operations such as closing opera-318 tions should be applied to the Tamp region for higher tamper 319 detection rate.

320
Tamp region (i, j) = W Ext_Auth (i, j) ⊕ W Cal_Auth (i, j) (4) 321 After locating the tampered regions, the recovery bits 322 are then used to recover the tampered regions. Due to a 323 series of mapping operations have been applied to increase 324 the security of the watermark information, we should 325 de-map the watermark by ACM and diagonal mapping. 326 On one hand, from the watermarked image, 6 bits of each 327 block that represent the coefficients of SPIHT are retrieved 328 and construct the 256 × 256 matrix W Ext_Rec1 . In the 329 stage of watermark generation and embedding, in order 330 to satisfy the capacity requirements, we put zeros into 331 the spare bits. Therefore, in order to obtain the initial 332 bitstream, we should resize the matrix and extract the 333 significant bits.

334
After that, we can apply the SPIHT decoding and recon-335 struct the image from SPIHT coefficients. On the other hand, 336 we apply the Stucki Kernel [26] algorithm to W Ext_Rec2 , and 337    To measure the quality of recovered images, the Peak Sig-352 nal to Noise Ratio (PSNR) [27] and Structural similarity 353 (SSIM) [28] are calculated using (8) and (10)    F1score, as shown in    obtain better quality of the watermarked images than the 445 existing TRLH [29]. Figure 9 shows the comparison results