Systematic Survey on Visually Meaningful Image Encryption Techniques

Due to advancements in technology, digital images are widely used in various applications like medical field, military communication, remote sensing, etc. These images may contain sensitive and confidential information. Therefore, images are required to be protected from unauthorized access. Many image protection techniques have been proposed in past years. The most common technique to protect the images is encryption. In this technique, a secret key and an encryption algorithm are used to change the plain image into an encrypted image. The encrypted image looks like a noisy image and can easily attract the attacker’s attention. If an image gets captured and stacked, sensitive information can be revealed. In this regard, Visually Meaningful Encrypted Image (VMEI) technique is developed, which initially encrypts the original image and then hides it into a reference image. The final encrypted image looks like a normal image. Hence, the VMEI technique provides more security as compared to simple image encryption techniques. Therefore, a systematic survey of existing VMEI techniques is presented in this paper. The VMEI techniques are divided into different categories based on their characteristics. Moreover, this paper elaborates and investigates the improvements and analyses performed on VMEI techniques based on various evaluation parameters. These evaluation parameters are divided into different categories such as security attacks, encryption key attacks, quality analysis, and noise attacks. Finally, this paper discusses the potential applications and future challenges of VMEI techniques.

tional information storage and transmission methods because 23 of fast access time, easy sharing, and many other advantages. 24 Information may be in the form of a text document, digital 25 image, audio, and video [1]. Information security is essential 26 to keep it secure from unauthorized access, release, disclo- 27 The associate editor coordinating the review of this manuscript and approving it for publication was Yang Liu . sure, alteration, copy, or damage, as it is an asset for any 28 organization or business [2]. 29 A digital image is one of the forms of data that is used 30 to store pictorial information and to communicate secret 31 data over the computer network [3]. Image encryption is 32 a cryptographic method to encode a secret image in such 33 a way that the unauthorized user can't understand it [4]. 34 In recent years, various image encryption algorithms are 35 developed. These algorithms are mainly based on spatial 36 and frequency domains. In the spatial domain, algorithms 37 usually jumble the pixels through matrix transformation and 38 extinguish the pixel's correlation of the secret image. In the 39 Here, C is the control parameter, with value lies in between 84 0 to 4. cs 0 and cs n represents the initial and generated random 85 value of chaotic map respectively [11]. In Algorithm 1, the 86 procedure to encrypt a M × N color image is illustrated. 87 To represent a grayscale image, one byte for each pixel is 88 required. One byte can store 0 to 255 values that cover all 89 possible shades of a gray color. A grayscale image is denoted 90 as a two-dimensional array of bytes in memory and this array 91 is called a channel. A grayscale image has only one channel. 92 For the color image (using RGB color space), 3 bytes are 93 required for each, one byte for each color. Hence, a color 94 image has three channels. The encryption process is required 95 for three matrices in the case of color image whereas, there is 96 a single matrix to be encrypted in a grayscale image. Hence, 97 the encryption burden is high for the color image as compared 98 to the grayscale image. However, the color image encryption 99 method could be complex because there are three matrices 100 that can be manipulated with each other that provide high 101 security. Moreover, we can use the same algorithm to encrypt 102 the color image that is designed for encrypting a grayscale 103 image by applying the same procedure to all three channels of 104 the color image that is designed for one channel of a grayscale 105 image. 106 In Algorithm 1, a secret color image I is converted into 107 one dimensional matrix of size M × 3N . The initial values 108 for the chaotic map and control parameter are the secret 109 keys. The generated chaotic map is sorted in ascending order 110 to get the sorted sequence S . Finally, to get the encrypted 111 noise-like image, the one dimensional I matrix is permuted 112 according to CS [12]. Algorithm 1 depicts the basic example 113 of chaotic map-based encryption. Similarly, the other hyper 114 chaotic map-based encryption techniques are implemented 115 [13], [14].

116
The second phase is embedding an encrypted image I into 117 reference image R. This embedding is performed in frequency 118 domain through Discrete wavelet transform (DWT) [15], 119 [16], [17]. In this, the reference image is decomposed into 120 four matrices denoted as C a (approximation matrix), C h , C v 121 and C d [5]. The parameter set P is used to define the filters 122 used in DWT. The DWT is defined as the following equation-123 VMEI = DWT I , R, P (2) 124 VOLUME 10, 2022 Algorithm 1 Pre-Encryption Input: Color image I of size M × N. Output: Encrypted Image I Step 1: The I is converted into one dimensional I = p 0 , p 1 , . . . . . . p M×3N matrix of size M × 3N Step 2: The initial values cs 0 and control parameter C are used as secret keys. The chaotic sequence CS is obtained using these keys Step 3: Sort the chaotic sequence to obtain CS = cs 0 , cs 1 , . . . . . .cs M ×3N Step 4: I pixels are permuted according to CS and I is obtained as-I (i) = I CS (i) Step 5: The obtained encrypted image I is noise-like

Algorithm 2 Embedding
Input: Encrypted Image I of size M × N , Reference image R of size 2M × 2N , and parameter P Output: VMEI Step 1: Decompose R into four matrices C a , C h , C v , and C d Step

End for End for
Step 3: Apply inverse DWT to C a , C h , C v , and C d In this paper, various problems associated with the existing 152 VMEI techniques are examined. The possible solutions to 153 these problems are also discussed.

154
To accomplish the aim of answering the research ques-155 tions, the study comprises databases of ACM, ScienceDirect, 156 IEEE, SpringerLink, and Google Scholar. These databases 157 are used to search and retrieve the published works on the 158 VMEI techniques. Figure 3 shows the number of research 159 papers reported over the last five years, while Figure 4 shows 160 the flowchart of the selection of articles for conducting a 161 systematic survey. The keyword used is visually meaningful 162 image encryption and the published articles' year range is 163 taken from 2015 to 2022. Based on this keyword entered in the above-mentioned 165 databases, a total of 22878 articles were found. By screening 166 duplicate articles and based on the title and abstract of the 167 articles, 22658 were sorted out. After the detailed evalua-168 tion, 228 articles were found irrelevant. The total number 169 of included articles is 78. Out of the 38 articles are VMEI 170 specific and included in the survey.

172
The key contributions of this paper are:  The bad visual quality of VMEI attracts the attacker and 224 represents the sign of encryption. Some researchers investi-225 gated the methods to improve the visual quality of VMEI by 226 improving the embedding technique. Kanso and Ghebleh [18] 227 improved the quality of VMEI and image security as com-228 pared to [5]. In this, a 3D chaotic map was used for encryption 229 in the first phase and in the second phase, a 2D Lifting 230 Wavelet Transform (LWT) was used to generate high quality 231 VMEI. The 3D chaotic map was utilized to increase the image 232 security, as these maps are highly sensitive to initial condi-233 tions and generate pseudorandom numbers [19]. LWT is an 234 improved implementation of DWT as it is fast in computation, 235 requires less memory, and produces a better-quality recov-236 ered image. In DWT, the wavelet coefficients are rounded 237 off to the whole number, so the lossless recovery is not 238 possible [20]. 239 Manikandan and Masilamani [21] proposed improvement 240 in the embedding technique by using DWT and Arnold 241 transform. A reference image was transformed into LL, LH, 242 HL, and HH sub-bands using DWT. An original image was 243 encrypted using an efficient algorithm and decomposed into 244 two matrices. Then, the Arnold transform was applied to each 245 matrix and put into LH and HH sub-bands of the reference 246 image. To determine the periodicity of the Arnold transforms, 247 the additional recovery image information was embedded in 248 the LL sub-band using the conventional LSB data hiding tech-249 nique. Arnold transform is based on the image scrambling 250 method. In this method, pixels are scrambled by the iterative 251 encoding process. The number of iterations is used as an 252 encryption key to recover the image [22]. 253 Yang et al. [23] used discrete quantum walks for the 254 key generation. This key is used for image encryption. 255 An encrypted image is divided into three parts using DWT 256 and embedded into the reference image. The implementation 257 of a quantum walks-based algorithm enhanced the VMEI 258 texture. Yang et al. [24] presented a method in which the 259 original image was encrypted using Qi hyper-chaotic system. 260 Block based discrete cosine transform (DCT) was applied to 261 get the coefficient matrix. After that, singular value decom-262 position (SVD) was applied to each DCT coefficient matrix. 263 This singular value was then embedded into the reference 264 image. A color reference image was converted from RGB to 265 YCbCr color space. DWT was performed on both Cb and Cr. 266 Then, block DCT was used to embed the data into a reference 267 image. This method contributed to the improvement of VMEI 268 quality. 269 Li et al. [25] presented a compressive-sensing-based data 270 hiding method in which a sparse representation of a reference 271 image was generated using a dictionary (such as a DCT  Armijo et al. [27] performed the embedding phase using the 279 2D integer haar wavelet transform and a substitution box.

280
In [28], chaotic maps are used for encryption and initial val- VMEI techniques [30], [31]. The salient regions of an 289 image contain useful information. Therefore, the detection 290 and encryption of these regions instead of the full image 291 reduce the computational complexity. In [30], the salient 292 features were detected and encrypted using a parametric 293 switching chaotic system-based image encryption algorithm. 294 Sun et al. [31] presented a method in which salient regions 295 were encrypted using a chaotic system and Deoxyribonucleic 296 Acid (DNA) coding. In both the methods [30] and [31], DWT 297 was used for embedding the encrypted image into a refer-298 ence image. Advantages of the DNA include enormous stor-299 age, massive parallelism, very low power consumption, and 300 high speed [32]. However, the implementation of the DNA 301 method requires expensive instruments and bio-molecular 302 laboratories [32].  There are mainly three steps followed in compressive sens-321 ing. First is the sparse representation of the matrix, the second 322 is measurement matrix generation which acts as an encryp-323 tion key, and the third is quantification or sparse recovery 324 process [33], which is done by converting the range of values 325 to a single value (see Figure 6). However, the measurement      The primary concern of VMEI techniques is to generate 446 secure encrypted images. The security level can be analyzed 447 by different parameters and these parameters are defined as 448 follows:

303
The degree of similarity between two contiguous pixels of 451 the image is evaluated using Correlation Coefficient Analysis 452 (CCA). It should be low, so that value of the neighbor pixel 453 cannot be estimated easily [62]. The correlation between two 454 adjacent pixel values x and y of an image can be calculated 455 as- covariance between x and y STD of x × STD of y (3) 457 Here, STD stands for standard deviation, which can be calcu-

463
Covariance between x and y is calculated as Information Entropy Analysis (IEA) evaluates the degree of 490 disorder or randomness of information in an image [23]. 491 When the value of information entropy is high then there 492 would be the least possibility of information leakage [64]. 493 Here, i represents the value of pixel and p(i) shows the prob-494 ability of i. The entropy of a given image can be calculated 495 as The value of IEA should be 8 ideally for a 256 (2 8 ) gray 498 level encrypted image [38]. The IEA values shown in Table 4 499 are the average values calculated for different images. The 500 ideal IEA value shows that all the pixels are distributed 501 randomly. However, it is generally less than 8 because in the 502 original image pixels are correlated. 503 Table 4 shows the CCA, HA, and IEA parameters which 504 are analyzed by the given references to show the effectiveness 505 of the existing VMEI methods. It can be observed that most 506 of the techniques have not evaluated all the parameters that 507 are necessary to stop the statistical attacks.

509
The encryption key used for image encryption is an impor-510 tant aspect of measuring the security level of the encryption 511 technique. There are two key analysis parameters such as key 512 space and key sensitivity. The key space represents the size of the key used by the 515 image encryption method. The size of the key must be large to 516 protest brute force attacks. It is an exhaustive task to try all the 517 possible combinations, for example, if the n-bit key is taken, 518 VOLUME 10, 2022 then 2n combinations are possible. For an effective image 519 encryption algorithm, key space must be at least 2100 or 520 larger than this [31]. To persist brute force attack it should be 521 large enough. Table 5 shows key space taken by given VMEI good key sensitivity if an original image cannot be recovered 528 even if there is a slight change in the key [31]. Table 5 shows 529 the VMEI techniques, which performed a key sensitivity test  Here, m represents the maximum value of the pixel. For 551 example, if the pixel value is represented in 8 bits, then 552 maximum value is 255. MSE stands for mean square error, 553 which is calculated as Here, R and C represent the rows and columns of the image 556 matrix. I 1 and I 2 represent the input images, which are being 557 compared. The values of PSNR for a VMEI and reconstructed image 559 should be at least 30 decibels or higher for an 8-bit image 560 and at least 60 decibels for the 16-bit image. The higher 561 PSNR value represents better the visual quality. Table 6 and 562 Figure 8 shows the VMEI techniques those evaluated the 563 PSNR of VMEI [18], [21], [23], [24], [26]

571
SSIM is used to assess luminance, contrast, and structure 572 comparisons between original and reconstructed images [65], 573 [66]. It is calculated among the same size windows of an 574 image. If x and y are two windows, then it is calculated as where c 1 and c 2 are the variables to stabilize the division with 577 weak denominator, 578 c 1 = (0.01, dr) 2 andc 2 = (0.03, dr) 2 (10) When an encrypted image is transmitted over the network 605 then because of some network fault or other faults, some 606 data may be lost. Hence, the quality of recovered image after 607 decryption may be affected [42]. The algorithm should be 608 capable of handling data loss and the quality of recovered 609 image must be unaffected.  Table 7 shows that most of the existing VMEI techniques 613 do not satisfy all the parameters. Therefore, the development 614 of an improved VMEI technique is still an open area. KPA and CPA are the types of cryptographic attacks. In KPA, 618 the attacker knows the original image and its corresponding 619 encrypted image. However, in CPA, the encryption method 620 is known to the attacker, hence attacker can encrypt any ran-621 domly chosen image and get the encrypted form of that image 622 [67]. These attacks are generally applied to find the original 623 image by cryptanalysis. An encryption method should be 624 effective to handle such types of attacks.  where data is stored or transferred over the computer network.

635
Some of the applications are as follows 636 1) Medical Applications: When a digital medical image is 637 exchanged among physicians over the network, the confiden-638 tiality of the image is an important factor to consider because 639 the image may contain sensitive data [11], [68].   other data are the main concern because it is stored on the 652 network [71]. images, text audio, video, etc.) is shared over the network. 656 It requires protection against unauthorized access, hacking, 657 & various types of attacks [72]. 658 6) Disaster Management: There are various uses of digital 659 images involved in real-time monitoring of natural disasters 660 like earthquakes, fire, floods, etc., to reach the disaster-struck 661 areas. The authorities need to protect these images from 662 authorized access [73]. 663 7) Telemedicine System: In modern technology, these sys-664 tems are used for medical consultation remotely. The digital 665 images are shared between doctors and patients like prescrip-666 tions, medical reports, etc., therefore it requires encryption 667 techniques to secure the images [55], [74]. 668 8) Remote Sensing: The process of monitoring and captur-669 ing information remotely, typically by satellites or aircraft is 670 called remote sensing. These images need to be protected to 671 maintain the privacy of data [75]. 672 9) Intelligent Transport System: The digital images are 673 captured through the cameras in the field of traffic and vehicle 674 management on roads to provide better services to the users. 675 It is necessary to secure these images to avoid misuse [76]. VMEI techniques perform encryption in two phases of 687 encryption and embedding, so the computational burden is 688 high. Hence, there is the requirement for a computationally 689 fast technique that could perform rapidly by using some speed 690 increasing technique like parallel processing.

692
In VMEI techniques, an encrypted image is implanted into 693 the cover image. Hence, the visual quality of the recovered 694 image is reduced. This necessitates betterment in the image 695 quality of the recovered image.

697
In the modern world, AI is one of the evolving technolo-698 gies. Implementation of AI techniques like a neural network, 699 genetic algorithm, and fuzzy logic, etc., in image processing, 700 requires wider investigation. It could be implemented in both 701 the phases of encryption and embedding to generate a VMEI. 702

703
The abilities of quantum computing are to store an enormous 704 amount of data, very high speed, and require less power as 705 compared to conventional computers. The computation speed 706 rithm based on ZigZag transform and LL compound chaotic system,'' Opt.