Received 23 August 2022, accepted 2 September 2022, date of publication 8 September 2022, date of current version 16 September 2022.

Digital Object Identifier 10.1109/ACCESS.2022.3204873

Fast Chaotic Image Encryption Algorithm Using a

Novel Divide and Conquer Diffusion Strategy
BIN GE 1, ZHIHUA SHEN1 , AND JINBAO ZHANG 2
1 College of Electronic Information Engineering, Nantong Vocational University, Nantong 226007, China
2 School of Information Science and Technology, Nantong University, Nantong 226019, China

Corresponding author: Bin Ge ([email protected])

This work was supported in part by the National Natural Science Foundation of China under Grant 62004108; in part by the Natural
Science Foundation for Universities of Jiangsu Province under Grant 21KJB520040, Grant 20KJD510001, and Grant 21KJB520041; in
part by the Qing Lan Project of Colleges and Universities in Jiangsu Province under Grant (2020) 10 ; and in part by the
Nantong Science and Technology Plan Project under Grant JC2021036 and Grant JC2021147.

1 ABSTRACT To protect image privacy in real-time transmission, a fast chaotic image encryption algorithm
2 using a novel divide and conquer diffusion strategy is proposed in this paper. Firstly, a fast pseudo-
3 random sequence generator is constructed using 2D hyperchaotic systems. And by performing bitwise
4 XOR operation between two integer sequences obtained from different systems, the final key stream used
5 for diffusion will have better randomness and unpredictability. Secondly, to achieve divide and conquer
6 diffusion, the plain image is divided into three parts, then a parallel CBC-based diffusion method can
7 simultaneously act on upper and bottom parts (or left and right parts), which achieves a time complexity
8 of O(W + H ). Moreover, within the diffusion process, the mechanism of information interaction on the
9 central row or column further enhances avalanche effect to resist differential attack. Thirdly, the session key
10 of the proposed algorithm is a mixture of the plain image’s hash value and a true random sequence, which
11 not only improve plaintext sensitivity but also realize one-time pad. Finally, experimental results indicate
12 the superiority of our algorithm to resist statistical, chosen-plaintext, entropy, and other common attacks.
13 Furthermore, by comparison with previous works, our algorithm preforms much faster execution speed with
14 only average of 0.08s to encrypt images of size 512 × 512, while it provide the same or even higher level
15 of security. Therefore, the proposed algorithm can meet security and efficiency requirements of real-time
16 communications of image data.

17 INDEX TERMS Image encryption, hyperchaotic system, fast encryption, divide and conquer diffusion,
18 one-time pad.

19 I. INTRODUCTION privacy, business, or even military information, the pri- 29

20 With the fast evolution of mobile devices and wireless net- vacy issue of digital images has attracted much concern 30

21 works, various types of multimedia data are produced and worldwide [6], [7], [8]. 31

22 transmitted on the Internet at a phenomenal rate [1], [2], [3]. As we all know, encryption techniques are always the most 32

23 As the main form of expression for multimedia, digital basic tool to protect data security. At present, although block 33

24 images are gaining popularity owing to their advantages of cipher is still the first choice to protect data privacy [9], [10], 34

25 vividness and understandability. However, the openness of it confronts fatal flaws when encrypting digital images. The 35

26 the Internet not only provides convenience but also makes high redundancy of image data will cause heavier time 36

27 it vulnerable to attackers [4], [5]. Since more and more consumption than textual data, let alone invalid encryption 37

28 organizations and individuals rely on images to exchange may occur now and then due to strong correlations between 38

pixels. Therefore, it is urgent to design specific encryption 39

The associate editor coordinating the review of this manuscript and algorithms for digital images [12], [13], [14], [15]. Since a 40

approving it for publication was Jeon Gwanggil . random permutation of pixel positions can quickly destroy the 41

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
95986 VOLUME 10, 2022
 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

42 readability of an image, many scrambling algorithms, such as process. Zhao et al. [44] segmented the plain image into mul- 98

43 Arnold transformation [16], Zigzag transformation [17], and tiple blocks, then all blocks were concurrently encrypted 99

44 magic cube transformation [18], have emerged. But unfortu- using a forward-then-backward diffusion strategy. But the 100

45 nately, an encrypted image only using scrambling is vulnera- lack of diffusion between blocks leaves a potential gap for 101

46 ble to attackers because it still keeps the statistical properties chosen-plaintext attacks. Wang and Zhao [45] applied the 102

47 of the plain image [19], [20]. Hence, a secure image encryp- CBC process on matrix rather than on sequence, thus pix- 103

48 tion algorithm must convert the image to a noise-like image els in a row/column could be encrypted in parallel, and by 104

49 not only at the visual level but also at the statistical level. performing a forward-then-reverse diffusion both in vertical 105

50 Although the application of DNA encoding [21], compressive and horizontal directions, a cipher image with sufficient dif- 106

51 sensing [22], hash function [23], and other techniques [24] fusion could fast output. However, the problem is that phase- 107

52 have promoted the progress of image encryption, the chaotic space reconstruction attacks may be a huge threat to this 108

53 system stands out for constructing a fast and secure cryptosys- algorithm due to the employed low-dimensional chaotic map. 109

54 tem due to its strong initial sensitivity and a long period of Liu et al. [40] not only implemented the CBC process on the 110

55 unpredictability [25], [26], [27], [28]. matrix but also further enhanced the complexity of diffusion 111

56 Since the first report of an algorithm using Logistic by embedding it with scrambling operation. But only one- 112

57 map [29] and the maturity of chaos control theory [30], more round diffusion can hardly guarantee this algorithm to resist 113

58 and more researchers have tried to solve image encryption chosen-plaintext attacks, while more diffusion round means 114

59 problems using chaotic systems. In 1998, Fridrich [31] first its higher time consumption. 115

60 proposed a general permutation-diffusion structure for chaos- Therefore, to solve the existing problems of the above- 116

61 based image encryption algorithms and stressed that the secu- mentioned CBC-based image encryption algorithms, this 117

62 rity of algorithm almost depended on the effect of diffusion, paper presents a fast chaotic image encryption algorithm 118

63 which has also been proven by many following cryptanalyses using a novel divide and conquer diffusion strategy. First, 119

64 [32], [33], [34]. Although modulo multiplication used in on basis of a plaintext-related session key calculated by a 120

65 block cipher has proven to be a feasible approach for suffi- secure hash function [46], [47], [48], we further mixt it with 121

66 cient diffusion [35], [36], the high time consumption limits its an external random sequence to obtain a one-time session 122

67 availability in image encryption. Therefore, Chen et al. [37], key. Then, multiple chaotic systems, including two 2D hyper- 123

68 Mao et al. [38] employed modulo addition and bitwise XOR chaotic systems and two cascade chaotic maps, are employed 124

69 (exclusive OR) to propose a more popular diffusion structure to generate key streams with better cryptographic properties. 125

70 based on CBC (cipher block chain). Thus, a pixel is not Afterward, a novel divide and conquer diffusion strategy is 126

71 only encrypted by a secret key but also affected by a previ- proposed to accelerate encryption speed with enough encryp- 127

72 ous cipher pixel. Furthermore, to enhance the capability of tion strength. Besides, a corresponding padding method is 128

73 resistance to differential attacks, many algorithms may apply also presented to meet the image size requirements of the 129

74 Chen’s diffusion process (or its variants) more than twice novel diffusion strategy. 130

75 by letting the last encrypted pixel of the previous diffusion The highlights of the present work are summarized below: 131

76 process be a new initial vector of the following diffusion [39]. (1) A fast chaotic image encryption algorithm using a 132

77 Recently, some researchers indicate that the separate exe- novel divide and conquer diffusion strategy is pro- 133

78 cution of permutation and diffusion (whether permutation- posed. Owing to its high security performance and fast 134

79 then-diffusion or diffusion-then-permutation) may lead to the execution speed, the proposed algorithm provides a 135

80 invalidation of the whole algorithm caused by the weakness competitive solution for real-time communications of 136

81 of any part. Hence, Liu et al. [40] attempted to apply diffusion massive image data. 137

82 operation during the row (or column) permutation. On the (2) We construct a fast cryptographic pseudo-random gen- 138

83 contrary, Li et al. [41] chose to embed pixel scrambling erator consisting of two 2D hyperchaotic systems. Due 139

84 operation into the diffusion process. to its discrete structure but strong chaos properties, the 140

85 However, along with the deepening of research, a growing 2D hyperchaotic system iterates much faster than a 141

86 body of evidence suggests that the CBC structure will also continuous hyperchaotic system but provides chaotic 142

87 meet efficiency bottlenecks [42], [43]. Typically, to mag- sequences with the same level of quality [49], [50]. 143

88 nify minor changes of any pixel to the whole cipher image, Furthermore, by performing bitwise XOR operations 144

89 an image with W columns and H rows is normally reshaped on quantified chaotic sequences, a more uniformly dis- 145

90 to a sequence with length up to W × H , then it will tributed key stream can be fast obtained, which can 146

91 be encrypted pixel by pixel from left to right [37], [38], also enhance its unpredictability and complexity to 147

92 [39], [41]. As we can see, the execution of traditional CBC- resist phase-space reconstruction attacks. Finally, the 148

93 based diffusion process is serial and only handles one pixel key stream will be reshaped to a W × H matrix (for an 149

94 at once, so most existing algorithms are unable to cope image with the size of W ×H ) to meet the requirements 150

95 with real-time secure communications of massive image data. of the divide and conquer diffusion strategy. 151

96 Hence, to overcome this defect, various types of parallel (3) A novel divide and conquer diffusion strategy is 152

97 techniques have been introduced to accelerate the diffusion designed to accelerate image encryption. The divide 153

VOLUME 10, 2022 95987

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

154 and conquer strategy usually breaks down a big prob- where h and γ are two control parameters of the GTDHCS. 203

155 lem into several sub-problems for a simultaneous solu- It can be seen from Figure 2 that when h = 5 and γ = 5, 204

156 tion, which can significantly improve computational the GTDHCS demonstrates hyper chaos properties since it 205

157 efficiency. In this paper, we attempt to divide the plain now contains two positive Lyapunov exponents. Therefore, 206

158 image into two parts for simultaneous diffusion to the GTDHCS can be utilized as another component of secure 207

159 achieve a time complexity of only O(W +H ), while the pseudo-random generator to produce diffusion keys. 208

160 central row/column acts a more important role. In phase

161 one, the plain image is firstly divided into upper part,
162 bottom part, and central row. Next, in the first round of
163 simultaneous diffusion on two parts, pixels are paral-
164 lel encrypted row by row. Then, after the initial vec-
165 tor is updated to the central row (it is encrypted by
166 both parts), the second round of simultaneous diffu-
167 sion can realize sufficient encryption in the vertical
168 direction. At last, after the image is further encrypted
169 by a two-round simultaneous diffusion on the left and
FIGURE 1. Lyapunov exponents spectrum of GTDHCS, (a) α = 2 and
170 right parts, we can obtain a fully encrypted image. β ∈ (0, 1], (b) α = (0, 2] and β = 1.
171 Moreover, the embedded permutation operation during
172 diffusion further enhances the encryption strength of
173 the proposed algorithm.
174 The remainder of this paper is organized as follows.
175 Section II introduces the employed chaotic systems and their
176 roles in this work. The details of the proposed image encryp-
177 tion algorithm are presented in Section III. In Section IV,
178 the applicability and superiority of the proposed image algo-
179 rithm are demonstrated by experimental results and analyses.
180 Finally, conclusions are drawn in Section V.

181 II. PRELIMINARY WORKS FIGURE 2. Lyapunov exponents spectrum of CTDHCS, (a) h = 5 and
182 The employed chaotic systems are presented in this section, γ ∈ (0, 5], (b) h ∈ (0, 5] and γ = 5.

183 and their roles in our algorithm are also elaborated.

184 A. TWO 2D HYPERCHAOTIC SYSTEMS B. TWO CASCADE CHAOTIC MAPS 209

185 Among various types of chaotic systems, the 2D hyperchaotic In 2021, by using a universal method, Yuan et al. [51] con- 210

186 system has the advantages of fast iteration speed and strong structed a series of cascade chaotic maps to provide better 211

187 robustness against degradation. Hence, this paper utilizes randomness and stronger robustness for image encryption. 212

188 two different 2D hyperchaotic systems to produce an unpre- To fast and sufficiently scramble pixel’s positions, our 213

189 dictable pseudo-random sequence with a long period and algorithm employs the cascade Logistic-Logistic chaotic map 214

190 good randomness. (CLLCM), which is given by 215

191 The cross 2D hyperchaotic system (CTDHCS) [49] is y1 (i + 1) = 1 − 2(1 − 2y1 (i) ) 2 2
(3) 216
192 defined as
α Meanwhile, our algorithm utilizes the cascade Logistic-
  
217
 x (i + 1) = sin
193
1
sin(x2 (i)) (1) Sine chaotic map (CLSCM), as shown in Equation (4), to pro- 218


x2 (i + 1) = β sin (π (x1 (i) + x2 (i))) duce an initial vector with high-quality for the CBC-based 219

diffusion process. 220

194 where α and β are two control parameters of the CTDHCS.
195 As shown in Figure 1, when α = 2 and β = 1, two positive y2 = 4 sin(πxi )(1 − sin(πxi )) (4) 221

196 Lyapunov exponents appear, indicating that the CTDHCS has As demonstrated in Figure 3, the largest Lyapunov expo- 222
197 evolved into a hyper chaos state. Hence, it can be used as nents of CLLCM and CLSCM are greater than the original 223
198 a component to construct a proper pseudo-random generator Logistic map or Sine map, indicating that they can provide a 224
199 used for diffusion. better random number for our algorithm. 225
200 Gao’s 2D hyperchaotic system (GTDHCS) [50] is
201 defined as III. THE PROPOSED IMAGE ENCRYPTION ALGORITHM 226
  
 x (i + 1) = sin hπ The proposed algorithm is detailed in this section, includ- 227
3
202 sin(x4 (i)) (2) ing 1) The generation method of a plaintext-related one- 228

x4 (i + 1) = γ sin(πx3 (i)x4 (i)) time session key; 2) The padding strategy of plain image;

229

95988 VOLUME 10, 2022

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

initial states of CTDHCS, x3 (1) and x4 (1) are initial states 257

of GTDHCS, y1 (1) and y2 (1) are initial states of CLLCM and 258

CLSCM respectively. 259

Step A-4: To overcome the security risks caused by tran- 260

sient effect, pre-iteration is necessary for the chaotic system 261

used in encryption algorithms. Since pre-iteration also affects 262

the encryption result, this paper handles the remaining part 263

of H by 264


 p1 = bitxor(hex2dec(H(58 : 60)),

hex2dec(H(49 : 51))) + 20





 p = bitxor(hex2dec(H(60 : 62)),
2
(6) 265


 hex2dec(H(52 : 54))) + 20

 p = bitxor(hex2dec(H(62 : 64)),
 3


hex2dec(H(55 : 57))) + 20


to obtain the pre-iteration p1 for CTDHCS and GTDHCS, 266

FIGURE 3. Largest Lyapunov exponents of (a) CLLCM, (b) CLSCM, p2 for CLLCM, and p3 for CLSCM. 267
(c) Logistic map, and (d) Sine map.
B. THE PADDING STRATEGY OF PLAIN IMAGE 268

To achieve the divide and conquer diffusion strategy within 269

230 3) The generation method of initial vector; 4) The gener- our encryption process, the plain image must be divided into 270

231 ation method of diffusion keys; 5) The generation method three parts: left half, central column, and right half when 271

232 of permutation keys; 6) The divide and conquer diffu- encryption in the horizontal direction; or upper half, bottom 272

233 sion based encryption process; 7) The decryption process. half, and central row when encryption in the vertical direc- 273

234 Figure 4 exhibits the flow diagram of encryption end. tion. It indicates that the plain image should have an odd 274

number of rows and columns. Since not all images meet this 275

235 A. THE GENERATION METHOD OF A PLAINTEXT-RELATED requirement, this paper applies Algorithm 1 to pad a W × H 276

236 ONE-TIME SESSION KEY image if necessary, and the details are as follows: 277

237 Establishing associations between the plain image and ses- C. THE GENERATION METHOD OF DIFFUSION KEYS 278
238 sion key is the most effective approach for enhancing the To obtain enough random keys for the proposed diffusion 279
239 plaintext sensitivity of a chaos-based encryption algorithm. strategy, this paper inputs the extracted parameters into CTD- 280
240 Moreover, this paper firstly performs SHA-256 on a plain HCS and GTDHCS and then employs Algorithm 2to con- 281
241 image to obtain its hash value, then further mixes the result struct a key matrix Kd with the same size as P. 282
242 with a true random sequence to obtain a plaintext-related one-
243 time session key, and ultimately extracts the initial values of D. THE GENERATION METHOD OF PERMUTATION KEYS 283

244 chaotic maps. The steps are detailed below: To realize high-strength encryption, simultaneous scrambling 284

245 Step A-1: Input the plain image into the SHA-256 function operations are performed during the diffusion process. Thus, 285

246 to generate its hash value H, which is a hexadecimal sequence we employ CLLCM to randomly control the circular shift of 286

247 with a length of 64. row and column, then the keys for scrambling are produced 287

248 Step A-2: Mix H with an external key EK (it consists of by the following steps: 288

249 64 true random hexadecimal numbers) using SK = bitxor Step D-1: Input y1 (1), p2 , H , and W into CLLCM. 289

250 (H, EK) to obtain the session key, where bitxor represents Step D-2: Iterate CLLCM p2 times, but drop the state value. 290

251 a bitwise exclusive-OR operation. Step D-3: Initialize an empty vector V. 291

252 Step A-3: Extract the initial values for the above-mentioned Step D-4: Continue to iterate CLLCM, and let V = {V, 292

253 chaotic systems by y1 (i)}, i = p2 +1, p2 +2, . . . , p2 +2H +2W −1, p2 +2H +2W . 293

 Step D-5: Obtain permutation keys Kpr and Kpc for rows 294


 x1 (1) = hex2dec(SK(1 : 12))/248 and columns by Equation (7), and then drop V. 295

x2 (1) = hex2dec(SK(10 : 21))/248


  j k
K pr1 = V (1 : H ) × 1015 %W

 

x3 (1) = hex2dec(SK(19 : 30))/248

 



(5)  j k
254
x4 (1) = hex2dec(SK(28 : 39))/248  K pc1 = V (H + 1 : H + W ) × 1015 %H





y1 (1) = hex2dec(SK(37 : 48))/248

 j k 296
15
K V


 
 pr2
 = (H + W + 1 : 2H + W ) × 10 %W
y2 (1) = hex2dec(SK(46 : 57))/248
 

 j k
 K pr2 = V (2H + W + 1 : 2H + 2W ) × 1015 %H


255 where the hex2dec operation can convert the hexadecimal
256 number to a decimal number, then x1 (1) and x2 (1) are (7) 297

VOLUME 10, 2022 95989

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

FIGURE 4. Flow diagram of encryption.

E. THE GENERATION METHOD OF INITIAL VECTOR 298

Algorithm 1 The Padding Strategy of Plain Image
An initial vector will also influence the effect of CBC-based 299
Input: The original plain image P
encryption. Therefore, we adopt on the following steps to fast 300
Output: The updated P, H, W after
create a random initial vector according to the size of the plain 301
padding
image in this paper: 302
1: [H , W ] ← size(P) Step E-1: Input y2 (1), p3, and H into CLSCM. 303
2: L ← {101,118,101,110,69,86,69,78} Step E-2: Iterate CLSCM p3 times, but drop the state value. 304
/∗ L is a label consisting of the decimal numbers of Step E-3: Initialize an empty vector V. 305
ASCII codes with e: 101, v: 118, n: 110, E: 69, V: 86, Step E-4: Continue to iterate CLSCM, and let V = {V, 306
and N: 78. Therefore, Lrepresents {E, V, E, N, e, v, e, n}, y2 (i)}, i = p3 + 1, p3 + 2, . . . , p3 + W -1, p3 + W . 307
which provides a basis for correctly removing padding Step E-5: Obtain initial vector I by I = bV × 1015 c%256, 308
after decryption. ∗/ then drop V. 309
3: if W %2 == 0
4: Initialize a vector V with a height of H with all ele- F. THE DIVIDE AND CONQUER DIFFUSION BASED 310
ments being zero ENCRYPTION PROCESS 311
5: V(1:8) ← LT // T represents the transpose operation Now, we can perform Algorithm 3 to encrypt a plain image, 312
6: V(H -7:H ) ← LT and Figure 5 illustrates how the pixel information is fast 313
7: P ← {P(:,1:W /2), V, P(:,W /2+1:W )} and sufficiently diffused to the whole cipher image by the 314
8: W ←W +1 proposed divide and conquer diffusion strategy. 315
9: Drop V Figure 5 demonstrates a demo of the encryption process 316
10: end if
consisting of two main stages: 1) Parallel encryption in the 317
11: if H %2 ==0
vertical direction. 2) Parallel encryption in the horizontal 318
12: Initialize a vector V with a width of W with all ele- direction. 319
ments being zero As illustrated in Figure 5 (a), in the first stage, the plain 320
13: V(1:8) ← L image is divided into the upper part, bottom part, and cen- 321
14: V(W -7:W ) ← L tral row, then the pixels are parallel encrypted row by row 322
15: P ← {P(1:H /2,:); V; P(H /2+1:H ,:)} from both upper and bottom parts. Meanwhile, diffusion is 323
16: H ←H +1 realized by applying the previous encrypted row to calculate 324
17: Drop V the ciphertext of the next row. Moreover, the complexity and 325
18: end if
security of the proposed encryption algorithm are further 326

95990 VOLUME 10, 2022

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

Algorithm 2 The Generation Method of Diffusion Keys Algorithm 3 The Encryption Process
Input: x1 (1), x2 (1), x3 (1), x4 (1), p1 , H , W Input: Kd , Kpr1 , Kpc1 , Kpr2 , Kpc2 , I, P
Output: Kd Output: C
//pre-iterations to eliminate the transient 1: Initialize a W × H Matrix TC with all elements being
effect zero
1: for i from 1 to p1 /∗Start diffusion in the vertical direction using I, Kd ,
2: x1 (i + 1) ← sin(α/ sin(x2 (i))) and Kpr1 simultaneously from the upper part and bottom
3: x2 (i + 1) ← β × sin(π × (x1 (i) + x2 (i))) part∗/
4: x3 (i + 1) ← sin(h × π/ sin(x4 (i))) 2: Temp1 ← (I + P(1, :))%256
5: x4 (i + 1) ← γ × sin(π × x3 (i) × x4 (i)) 3: Temp2 ← (I + P(H , :))%256
4: Temp1 ← bitxor(Kd (1, :), Temp1 )
6: end for
5: Temp2 ← bitxor(Kd (H , :), Temp2 )
//update x1 (1), x2 (1), x3 (1), x4 (1)
6: TC(1, :) ← circshift(Temp1 , Kpr1 (1))
7: x1 (1) ← x1 (p1 +1), x2 (1) ← x2 (p1 +1),
7: TC(H , :) ← circshift( Temp2 , Kpr1 (H ))
x3 (1) ← x3 (p1 +1), x4 (1) ← x4 (p1 +1) /∗Continue to diffuse the remaining rows simultaneously
Initialize two empty vectors V1 , V2 from the upper part and bottom part∗/
8: for i from 1 to dW ×H/2e-1
8: for i from 2 to (H -1)/2
9: x1 (i + 1) ← sin(α/ sin(x2 (i))) 9: Temp1 ← (TC(i-1, :) + P(i, :))%256
10: x2 (i + 1) ← β × sin(π × (x1 (i) + x2 (i))) 10: Temp2 ← (TC(H + 2-i, :) + P(H + 1-i, :))%256
11: x3 (i + 1) ← sin(h × π/ sin(x4 (i))) 11: Temp1 ← bitxor(Kd (1, :), Temp1 )
12: x4 (i + 1) ← γ × sin(π × x3 (i) × x4 (i)) 12: Temp2 ← bitxor( Kd (H +1-i, :), Temp2 )
13: V1 ← {V1 , x1 (i), x2 (i)} 13: TC(i, :) ← circshift(Temp1 , Kpr1 (i))
14: V2 ← {V2 , x4 (i), x3 (i)} 14: TC(H + 1-i, :) ← circshift( Temp2 , Kpr1 (H +1-i))
15: end for 15: end for
16: Initialize empty vectors T1 , T2 //Information interaction on the central row
17: T1 ← b(|V1 (1:W ×H )| - b|V1 (1:W ×H )|c)×1015 c%256 16: Temp ← (TC((H -1)/2, :)+ TC((H -1)/2+2, :)+P((H -
1)/2+1, :))%256
18: T2 ← b(|V1 (1:W ×H )| - b|V1 (1:W ×H )|c)×1015 c%256 17: Temp ← bitxor(Kd ((H -1)/2+1, :), Temp)
18: TC((H -1)/2+1, :) ← circshift(Temp, Kpr1 ((H -1)/2+1))
19: Drop V1 , V2 and then initialize an empty vector V /∗Repeat the above diffusion steps using Kd , Kpr2 , and
20: V ← bitxor(T1 , T2 ) the updated I
21: Kd ← reshape(V, H , W ) then drop V 19: I ← TC((H -1)/2+1, :)
20: Temp1 ← (I + TC(1, :))%256
21: Temp2 ← (I + TC(H , :))%256
21: Temp1 ← bitxor(Kd (1, :), Temp1 )
327 enhanced by interleaving the random row circular shift in the 22: Temp2 ← bitxor( Kd (H , :), Temp2 )
328 diffusion process. The central row plays an important role in 23: TC(1, :) ← circshift(Temp1 , Kpr2 (1))
329 information interaction between the upper part and bottom 24: TC(H , :) ← circshift( Temp2 , Kpr2 (H ))
330 part. By encrypting the central row using the ciphertext of 25: for i from 2 to (H -1)/2
331 both the upper and bottom parts and taking it as a new initial 26: Temp1 ← (TC(i-1, :) + TC(i, :))%256
332 vector, diffusion between the upper part and bottom part can 27: Temp2 ← (TC(H + 2-i, :) + TC(H + 1-i, :))%256
333 be implemented to finish sufficient encryption in the vertical 28: Temp1 ← bitxor(Kd (1, :), Temp1 )
334 direction. 29: Temp2 ← bitxor( Kd (H +1-i, :), Temp2 )
335 Next, as shown in Figure 5(b), to meet the requirement of 30: TC(i, :) ← circshift(Temp1 , Kpr2 (i))
336 avalanche effect, parallel encryption in the horizontal direc- 31: TC(H + 1-i, :) ← circshift( Temp2 , Kpr2 (H +1-i))
337 tion is essential to make the whole cipher image influenced by 32: end for
338 minor changes of a pixel. Thus, the temporary cipher image //Information interaction on central row
339 created by the first stage is divided into the left part, right part, 33: Temp ← (TC((H -1)/2, :)+ TC((H -1)/2+2, :)+TC((H -
340 and central column. Firstly, the central column is used as the 1)/2+1, :))%256
341 initial vector to connect the first stage and the second stage. 34: Temp ← bitxor(Kd ((H -1)/2+1, :), Temp)
342 Then, the pixels are parallel encrypted column by column 35: TC((H -1)/2+1, :) ← circshift(Temp, Kpr2 ((H -1)/2+1))
343 from both the left and right parts. Similar to the first stage, the /∗Start diffusion in the horizontal direction using Kd ,
344 central column plays an important role in achieving sufficient Kpc1 , and the updated I simultaneously from the left part
345 encryption in the vertical direction. and right part∗/

VOLUME 10, 2022 95991

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

Algorithm 3 (Continued.) The Encryption Process Algorithm 4 The Removing Padding Strategy of Decrypted
1: I ← TC(:, (W -1)/2+1) Image
2: Temp1 ← (I + TC(:, 1))%256 Input: The decrypted image D
3: Temp2 ← (I + TC(:, H ))%256 Output: The original plain image P
4: Temp1 ← bitxor(Kd (:, 1), Temp1 ) 1: [H , W ] ← size(D)
5: Temp2 ← bitxor( Kd (:, H ), Temp2 ) 2: L← {101,118,101,110,69,86,69,78}
6: TC(:, 1) ← circshift(Temp1 , Kpr1 (1)) // Remove the inserted row if needed
7: TC(:, H ) ← circshift( Temp2 , Kpr1 (H )) 3: ifbitand(D((H -1)/2+1, 1:8), D((H -1)/2+1, W -7:W ))
/∗Continue to diffuse the remaining rows simultaneously == L
from the left part and right part∗/ 4: D0 ← {D(1:(H -1)/2, :); D((H -1)/2+2:H , :)}
8: for i from 2 to (W -1)/2 5: H ←H -1
9: Temp1 ← (TC(:, i-1) + TC(:, i))%256 6: end if
10: Temp2 ← (TC(:, W + 2-i)+ TC(:, W + 1-i))%256 // Remove the inserted column if needed
11: Temp1 ← bitxor(Kd (:, 1), Temp1 ) 7: ifbitand(1:8, D0 ((W -1)/2+1), D0 (H -7:H , (W -1)/2+1))
12: Temp2 ← bitxor( Kd (:, W + 1-i), Temp2 ) == LT
13: TC(:, i) ← circshift(Temp1 , Kpc1 (i)) 8: P ← {D0 (:, 1:(W -1)/2), D0 (:,(W -1)/2+2):W }
14: TC(:, W + 1-i) ← circshift( Temp2 , Kpc1 (W +1-i)) 9: W ←W -1
15: end for 10: end if
//Information interaction on central column
16: Temp ← (TC(:, (W -1)/2)+TC(:, (W -1)/2+2)+TC(:, (H -
1)/2+1))%256
G. THE DECRYPTION PROCESS 350
17: Temp ← bitxor(:, Kd ((W -1)/2+1), Temp)
Since gray values are in the range of 0-255, all operations in 351
18: TC(:, (W -1)/2+1) ← circshift(Temp, Kpc1 ((W -1)/2+1))
Algorithm 3 have valid inverse operations. Hence, as shown 352
/∗Repeat the above diffusion steps using Kd , Kpc2 , and
in Figure 6, the original plain image can be recovered at 353
the updated I
decryption end using a correct session key. 354
19: I ← TC(:, (W -1)/2+1)
However, due to the padding strategy before the encryption 355
20: Temp1 ← (I + TC(:, 1))%256
process, the decrypted image is not always the original plain 356
21: Temp2 ← (I + TC(:, H ))%256
image. Therefore, Algorithm 4 must be applied to remove 357
22: Temp1 ← bitxor(Kd (:, 1), Temp1 )
the central row or column based on the labels inserted by 358
23: Temp2 ← bitxor( Kd (:, H ), Temp2 )
padding. 359
24: TC(:, 1) ← circshift(Temp1 , Kpr1 (1))
25: TC(:, H ) ← circshift( Temp2 , Kpr1 (H ))
26: for i from 2 to (W -1)/2
IV. EXPERIMENTAL RESULTS AND ANALYSES 360

27: Temp1 ← (TC(:, i-1) + TC(:, i))%256 The applicability of the proposed algorithm is evaluated by 361

28: Temp2 ← (TC(:, W + 2-i)+ TC(:, W + 1-i))%256 different kinds of tests in this section. Indeed, the analy- 362

29: Temp1 ← bitxor(Kd (:, 1), Temp1 ) ses and comparisons can demonstrate the superior security 363

30: Temp2 ← bitxor( Kd (:, W + 1-i), Temp2 ) and efficiency of our algorithm. All tests are performed on 364

31: TC(:, i) ← circshift(Temp1 , Kpc1 (i)) the MATLAB R2016b platform using a personal computer 365

32: TC(:, W + 1-i) ← circshift( Temp2 , Kpc1 (W +1-i)) equipped with Intel(R) Core (TM) i5 6500 CPU @ 3.20GHz, 366

33: end for

16 GB 2400Hz DDR4, 256 GB M.2 SSD, and running a 367

//Information interaction on central column 64-bit Window 7 operating system (professional edition). 368

34: Temp ← (TC(:, (W -1)/2)+TC(:, (W -1)/2+2)+TC(:, (H -

For simulation, we access Random.org [52] to obtain a true 369

1)/2+1))%256 random external key EK = {4c9df03c6c4ca7ead0b00b62069 370

35: Temp ← bitxor(:, Kd ((W -1)/2+1), Temp)

bde031799dcc44ee24bc572ff49b931c26253}. 371

36: TC(:, (W -1)/2+1) ← circshift(Temp, Kpc2 ((W -1)/ Several standard 8-bit gray images of the USC-SIPI image 372

2+1)) database are used as test objects: 1) ‘Lena’ of size 256 × 256; 373

37: C ← TC//Finally the cipher image C is obtained

2) ‘Resolution’ of size 256 × 256; 3) ‘Baboon’ of size 374

512 × 512; 4) ‘Ruler’ of size 512 × 512; 5) ‘Airplane’ of size 375

1024×1024; 6) ‘Male’ of size 1024×1024. Their encryption 376

and decryption results are presented in Figure 7. 377

The comparison of Figure 7 (column one) and Figure 7 378

346 In general, the proposed divide and conquer diffusion strat- (column two) indicates that all plain images containing rich 379

347 egy achieves high security by full encryption in both vertical information are successfully encrypted into totally disordered 380

348 and horizontal directions, while it provides a high operating images. Then, as shown in Figure 7 (column three), all cipher 381

349 efficiency with a time complexity of only O(W + H ). images can be recovered to original images by the decryption 382

95992 VOLUME 10, 2022

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

FIGURE 5. Demo of the encryption process. The effect of diffusion (a) in the vertical direction, and (b) in the horizontal
direction.

383 process using their corresponding session keys, which con- different and rich information. Then, in the fourth column, all 392

384 firm the validity of the proposed algorithm. the histograms are approximately flat, showing that our algo- 393

rithm can successfully hide the distributions of the original 394

385 A. HISTOGRAM ANALYSIS images. 395

386 The histogram is an important visual tool for observing the Furthermore, based on the histogram, the chi-square (χ 2 ) 396

387 distribution of an image’s pixels. A valid image encryption test [40] defined in Equation (8) is a vital tool for deep 397

388 algorithm should output a uniformly distributed cipher image quantitative analysis of the flatness of the histogram. 398

389 to resist frequency attacks. Figure 8 demonstrates the results L

X (fi − pi )2
390 of histogram tests. The fluctuant histograms of plain images χ2 = (8) 399

in Figure 8 (column two) indicate that each image contains pi

391 i=1

VOLUME 10, 2022 95993

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

FIGURE 6. Flow diagram of the decryption.

400 where L represents the grey level (for an 8-bit grayscale is to minimize such inner correlation. The lower the correla- 428

401 image, its maximum level L = 256), pi and fi are respectively tion, the higher the security. 429

402 the expected proportion and real proportion of each pixel For this test, 10,000 pairs of adjacent pixels in all direc- 430

403 value in a test image. For a uniform distributed cipher image, tions are randomly extracted from ‘Baboon’. By comparing 431

404 it should have an expected χ 2 = 293.24783 at the signifi- the scatter diagrams of the plain and cipher images illus- 432

405 cance level α = 0.05. Since all the χ 2 values of the above trated in Figure 9, the highly concentrated distributions all 433

406 cipher images in Table 1 are less than 293.24783, the corre- become random distributions in any direction after ‘Baboon’ 434

407 sponding P-values are far more than 0.05. The results indicate is encrypted by the proposed algorithm. 435

408 that our algorithm can achieve the goal to protect images with The indicator of γxy (correlation coefficient) provides an 436

409 any distribution against frequency analysis attacks. approach for quantitatively analyzing the degree of correla- 437

tion between adjacent pixels, which defined as 438

 N N
410 B. RANDOMNESS ANALYSIS X X
/N , yi /N

x̄ = x ȳ =

i

However, to avoid information leaking, the encrypted image

411


 i=1 i=1
should be random on both the visual side and the statistics

412 N

 P
413 side. Therefore, a further statistical test is conducted using (xi − x̄)(yi − ȳ) (9) 439

414 the SP 800-22 randomness test suit powered by NIST [53]. γ = s

 i=1

 xy N  N 
Meanwhile, the quality of the keys used in encryption is also

415

2 2
 P P
416 tested. Then, after all the objects under test are transformed


 (x i − x̄) (y i − ȳ)
i=1 i=1
417 into a binary stream, their P-values (with a significance level
418 α = 0.01) are obtained and listed in Table 2. Since all where xi and yi are two adjacent pixels in an image. Then, 440

419 the results are more than 0.01, the keys and cipher images after N is set to10,000, the results are shown in Table 3. 441

420 are random in all aspects with a confidence of 99%. Hence, Based on ‘Baboon’, Table 4 compares γxy of this work 442

421 the proposed algorithm can resist attacks based on statistical and other works. Therefore, the lower correlation coefficients 443

422 analysis. between adjacent pixels of this work indicate better resistance 444

against statistical attacks. 445

423 C. ANALYSIS OF CORRELATION BETWEEN D. INFORMATION ENTROPY ANALYSIS 446

424 ADJACENT PIXELS Information entropy is an essential measurement to judge the 447

425 Unlike text data, the strong inner correlation and high redun- disorder level of images, and a higher entropy indicates a 448

426 dancy make it hard to completely hide the pattern of a plain more random image. As given by Equation (10), the global 449

427 image. One of the most important targets of image encryption entropy H (g) can measure the global uncertainty of an image. 450

95994 VOLUME 10, 2022

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

FIGURE 7. Results of encryption and decryption. Plain images locate in the first column, while
their corresponding cipher images and decrypted images locate in the second and third
columns, respectively.

451 Here, gi represents a pixel value (for an 8-bit grayscale image, In an ideal cipher image, all pixel values are equiprobable, 455

452 L is a integer between 0 and 255), and p(gi ) expresses the indicating that the expected maximum H (g) = 8. Since the 456

453 proportion of each pixel value. results of entropy tests listed in Table 5 are all very close to 8, 457

X L it is confirmed that our algorithm achieves equivalent or even 458

454 H (g) = − p(gi ) log2 p(gi ) (10) better performance in resisting global entropy analysis-based 459

i=0 attacks. 460

VOLUME 10, 2022 95995

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

FIGURE 8. Results of histogram test. The second column presents histograms of plain images in column one. The fourth column presents
histograms of cipher images in column three.

95996 VOLUME 10, 2022

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

TABLE 1. Results of chi-square test.

TABLE 2. Results of randomness tests.

FIGURE 9. Scatter diagrams of the plain image in (a) the vertical direction, (b) the horizontal direction, and (c) the diagonal direction. The scatters of the
cipher image in (d) the vertical direction, (e) the horizontal direction, and (f) the diagonal direction.

461 In 2013, a more rigorous test called local entropy was where Si (i = 1, 2, . . . , Sk ) are k non-overlapping sub-images 465

462 proposed [54], which could detect the limitations of global containing Tb pixels randomly extracted from an image. The 466

463 entropy. Its equation is expressed by ideal value of the local entropy should be 7.902469317 (set 467

k (k, Tb ) = (30, 1936) and α = 0.05), but generally, if the 468

X H (Si )
464 H̄k,Tb (S) = (11) results fall into the range from 7.901901305 to 7.903037329, 469
k the tests could be considered passed. As listed in Table 6,
i=1 470

VOLUME 10, 2022 95997

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

TABLE 3. Results of correlation coefficients.

TABLE 4. Comparisons between different algorithms.

TABLE 5. Results of global entropy test.

471 since all local entropies of cipher images fall into the accep- calculates NPCR and UACI between encryption results. After 489

472 tance interval, it is confirmed that the proposed image encryp- the above test is repeated 100 times, the average values are 490

473 tion algorithm provides enough security to resist strong listed in Table 7. By comparison with the expected values 491

474 attacks of local entropy analysis. of other works, it is confirmed that the proposed algorithm 492

can provide enough security even facing strong differential 493

475 E. PLAINTEXT SENSITIVITY ANALYSIS attacks. 494

476 As mentioned earlier, a secure encryption algorithm should

477 output two completely different cipher images when two F. KEY SENSITIVITY ANALYSIS 495

478 plain images with a slight distinction are input. This guar- In practical applications, all details of an encryption algo- 496

479 antees the capability of a cryptosystem to resist strong dif- rithm are overt except the session key. This indicates that the 497

480 ferential attacks. In general, NPCR and UACI are two key privacy protection of image data depends on the quality of the 498

481 indicators to represent the degree of difference between the session key. For this purpose, the cipher image must be highly 499

482 two images. Suppose C1 and C2 are two different cipher sensitive to even slight modifications of the session key in 500

483 images, and D(i, j) =0 if C1 (i, j) =C2 (i, j) or D(i, j) =1 if two aspects: 1) Tiny changes of the session key will lead to 501

484 C1 (i, j) 6 =C2 (i, j). Then, NPCR and UACI can be defined as a significant difference between outputs when encrypting the 502

 same plain image; 2) Subtle perturbations of the session key 503

H X W
 X will result in a complete failure of decryption. 504
NPCR = D(i, j)/(HW ) × 100%


Using ‘Ruler’ as the test image, its corresponding session

505


i=1 j=1
485
H X W key SK = {5D64E5EC976C6C830C2555BCBE9B03294DC 506
 X |C 1 (i, j) − C 2 (i, j)| 5A94B8346579082C2B73EF5AD1643} is obtained via
/(HW ) × 100% 507


 UACI =

 255 Step A-2. Then, following Step A-3 and Step A-4, SK will be 508
i=1 j=1
486 (12) split into nine parts to extract the initial conditions of chaotic 509

systems. 510

487 For strict evaluation, this paper randomly picks one pixel In phase one, the least significant digit of nine parts is 511

488 from a test image, modifies its least significant digit, and then successively modified, and NPCR and UACI are calculated 512

95998 VOLUME 10, 2022

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

TABLE 6. Results of local entropy test.

TABLE 7. Average values of NPCR and UACI test (Unit: %).

513 between cipher images. As shown in Table 8, all values of Figure 11 and Figure 12 show the results of histogram 550

514 NPCR and UACI are close to 99.6094% and 33.4635%, indi- tests and scatter tests, while the other experimental results 551

515 cating that minor changes in the session key can significantly are listed in Table 10. It can be seen from Figure 11 and 552

516 affect the cipher image when encryption. Thus, the proposed Figure 12 that the ‘Black’ image and ‘White’ image are 553

517 algorithm is highly sensitive to minor modifications of the successfully encrypted to random images at the visual level. 554

518 encryption key. Then, as illustrated in Table 10, all the other results meet the 555

519 In phase two, the correct session key and modified ses- requirements of passing tests, which further indicates that our 556

520 sion key are used to decrypt the cipher image, respectively. algorithm can resist CPA. 557

521 As illustrated in Figure 10, only the correct session key can
522 recover the plain image while other decrypted images are still I. TIME COMPLEXITY ANALYSIS 558

523 disordered. The previous sections have confirmed that the proposed 559

524 Then, NPCR and UACI are used to further evaluate the image encryption algorithm has high security to resist com- 560

525 degree of the difference between the cipher image and the mon attacks. However, for real-time communication appli- 561

526 above decrypted images. The results listed in Table 9 show cations, fast encryption speed is also indispensable. Hence, 562

527 that tiny errors in the session key will cause the failure of the time complexity of the proposed algorithm and its actual 563

528 decryption without leaking any useful information. There- execution time on the Matlab platform are analyzed. Since 564

529 fore, our algorithm also has a high sensitivity to the decryp- the session key could be obtained before encrypting images, 565

530 tion key. and the time consumption of padding or removing padding of 566

images is quite negligible, this paper only discusses the time 567

531 G. KEY SPACE ANALYSIS consumption of other parts of the proposed image encryption 568

532 Besides the high sensitivity to key changes, a large enough algorithm. 569

533 key space of the session key is also essential to enabling For the generation method of diffusion keys, referring to 570

534 an encryption algorithm to resist brute-force attacks. In this Algorithm 2, its time complexity is O(WH/2). The gener- 571

535 paper, the 256-bit session key is obtained from the mixture ation method of permutation keys, referring to Step D-1 to 572

536 of a true random sequence and a unique identification of a Step D-5, has a time complexity of O(2W + 2H ). The 573

537 plain image, which ensures the one-time pad and randomness. generation method of initial vector, referring to Step E-1 to 574

538 In the above section, it has been proven that all bits of SK are Step E-5, has a complexity of O(W ). The encryption process, 575

539 valid to affect the output of encryption and decryption. Then, referring to Algorithm 3, has a time complexity of O(W +H ). 576

540 our algorithm can provide 2256 different session keys, which Then, our algorithm realizes a time complexity of O(WH/2+ 577

541 is much bigger than a secure threshold of 2128 to resist key 4W + 3H ). Attributed to the use of the novel divide and 578

542 exhaustion attacks [40], [55]. conquer diffusion strategy, the efficiency of our algorithm is 579

superior to that of other CBC-based works. 580

543 H. CHOSEN-PLAINTEXT ATTACKS ANALYSIS Then, the actual execution time of the proposed algorithm 581

544 The CPA (chosen-plaintext attack) is always a top threat to an is listed in Table 11. Even considering the low efficiency of 582

545 applicable encryption algorithm. Commonly, ‘Black’ (with MATLAB, our algorithm still achieves satisfactory results 583

546 all pixels being 0) and ‘White’ (with all pixels being 255) in comparison with other works. This is attributed to the 584

547 are two often used images to implement CPA. Hence, this optimization of the encryption process and the faster itera- 585

548 paper performs all the above security tests to verify whether tion speed of 2D hyperchaotic systems than high-order con- 586

549 the proposed algorithm could resist CPA. tinuous hyperchaotic systems. Combined with the previous 587

VOLUME 10, 2022 95999

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

TABLE 8. Results of the encryption key sensitivity test (Unit: %).

FIGURE 10. Decryption results from (a) the encrypted ‘Ruler’ with (b) a correct session key, and an incorrect
session key by changing (c) SK (12) = D, (d) SK (21) = 4, (e) SK (30) = 2, (f) SK (39) = 5, (g) SK (48) = 1,
(h) SK (57) = E , (i) SK (60) = C, (j) SK (62) = 7, (k) SK (64) = 2.

TABLE 9. Results of the decryption key sensitivity test (Unit: %).

588 outstanding performance of security tests, it can be concluded J. ROBUSTNESS ANALYSIS 591

589 that the proposed algorithm is suitable for massive image data During data transmission, data loss and noise pollution are 592

590 real-time communications. often unavoidable. However, due to its high redundancy, 593

96000 VOLUME 10, 2022

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

FIGURE 11. (a) ‘Black’ image, (b) its cipher image, and (c) the histogram of (b). (d) ‘White’ image,
(e) its cipher image, and (f) the histogram of (e).

FIGURE 12. Scatter diagrams of (a) the encrypted ‘Black’ image and (e) ‘White’ image in the vertical direction, horizontal direction, and diagonal direction.

TABLE 10. Security performance of our algorithm on ‘Black’ image and ‘White’ image.

594 minor errors do not affect the overall intelligibility of image Figure 13 demonstrates the decryption results of cipher 598

595 data. Hence, a useful image cryptosystem should have the images with data loss of varying degrees during transmission, 599

596 capability to recover a cipher image polluted by noise or which indicates that even though half the pixels are dis- 600

597 encountering data loss. carded, our algorithm can still recover it to a readable image. 601

VOLUME 10, 2022 96001

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

TABLE 11. Results of execution times (Unit: s).

FIGURE 13. Decryption results of cipher images with data lost (a) 1/16, (b) 1/8, (c) 1/4, and (d) 1/2.

FIGURE 14. Decryption results of cipher images with data lost 1/8 in different locations.

602 Figure 14 shows the decrypted images with a data loss of algorithm has certain robustness to resist noise pollution in 611

603 1/8 in different locations during transmission, meaning the low-quality channels. 612

604 location of data loss does not affect the decryption. Hence, Moreover, this paper utilizes PSNR test to check the quality 613

605 our algorithm has strong robustness to data loss in complex of the above decrypted images, which defined as 614

606 open channels.

607 Figure 15 shows the decryption results of cipher images 255 × 255
608 polluted by common noise of different intensities during PSNR = 10 lg (13) 615
W P
H
transmission. Since some important information can still (R(i, j) − D(i, j))2 /(WH )
P
609

610 be recognized to confirm ‘Baboon’ images, the proposed i=1 j=1

96002 VOLUME 10, 2022

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

FIGURE 15. Decryption results of cipher images polluted by Gaussian noise with an intensity of
(a) 0.001, (b) 0.003, (c) 0.005, and polluted by Salt&Peppers noise with an intensity of (d) 0.01,
(e) 0.03, (f) 0.05.

TABLE 12. Results of PSNR test.

616 where R represents the reference image and D represents the we must repeat the previous diffusion process by acting the 641

617 test image with data loss or pollution. As demonstrated in central row (which is already influenced by both upper and 642

618 Table 12, for the above decrypted images, all results of the bottom parts) as a new initial vector. Next, the temporary 643

619 PSNR test exceed 10dB, which further proves the proposed cipher image will be divided into three parts (left part, right 644

620 algorithm can provide privacy protection properly even on an part, and central column), thus a similar diffusion process as 645

621 unstable channel. phase one but in horizontal direction (here, it is launched by 646

the central column of the temporary cipher image) will be 647

622 V. CONCLUSION performed to achieve full encryption, which guarantees that 648

623 This paper presents a fast chaotic image encryption algorithm any minor changes in the plain image can lead to significant 649

624 using a novel divide and conquer diffusion strategy. First, for changes in the cipher image. Moreover, within the diffu- 650

625 any plain image, a 256-bit one-time session key is obtained sion process, CLLCM is employed to randomly scramble 651

626 from a mixture of its hash values and a true random sequence. pixel positions to further enhance the security of the pro- 652

627 Second, a pseudo-random sequence generator is designed posed algorithm. In addition, to make our algorithm adapt to 653

628 to produce diffusion keys. The employed 2D hyperchaotic images of any size, a padding strategy and its corresponding 654

629 systems, CTDHC and GTDHCS, overcome the high time removing strategy are also presented. Experimental results 655

630 consumption of solving continuous hyperchaotic systems. and comparisons indicate that the proposed algorithm can 656

631 In addition, the bitwise XOR operation between quantified provide more outstanding security performance to resist com- 657

632 hyperchaotic sequences are performed to further enhance mon attacks such as frequency analysis, differential analysis, 658

633 the statistical properties and anti-attack performance of the and statistical analysis because its P-values of χ 2 tests are 659

634 final key stream for diffusion. Then, we present a novel much larger than 0.05, P-values of NIST SP-800 22 tests are 660

635 divide and conquer diffusion strategy to accelerate the CBC- much larger than 0.01, correlation coefficients are closer to 661

636 based encryption scheme. In phase one, the plain image is 0, global entropies are closer to 8, and values of NPCR and 662

637 divided into three parts, thus a simultaneous diffusion can UACI tests are approaching ideal values. More importantly, 663

638 work on upper and bottom parts. In particular, the encryp- due to the time complexity of O(W + H ) of our encryption 664

639 tion is launched by an initial vector provided by CLSCM. process, our algorithm only costs 0.02s, 0.08s, and 0.31s 665

640 Then, to implement sufficient diffusion in vertical direction, to encrypt images of sizes 256 × 256, 512 × 512, and 666

VOLUME 10, 2022 96003

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

667 1024 × 1024, which meets the high efficiency requirements [21] H. Li, T. Li, W. Feng, J. Zhang, J. Zhang, L. Gan, and C. Li, ‘‘A novel 737

668 of real-time transmission. Overall, with continuous research image encryption scheme based on non-adjacent parallelable permutation 738
and dynamic DNA-level two-way diffusion,’’ J. Inf. Secur. Appl., vol. 61, 739
669 in chaos theory and parallel techniques, our algorithm can Sep. 2021, Art. no. 102844, doi: 10.1016/j.jisa.2021.102844. 740
670 provide security solutions for image data communication [22] X. Chai, Y. Tian, Z. Gan, Y. Lu, X. J. Wu, and G. Long, ‘‘A robust 741

671 scenarios. compressed sensing image encryption algorithm based on GAN 742
and CNN,’’ J. Mod. Opt., vol. 69, no. 2, pp. 103–120, 2022, doi: 743
10.1080/09500340.2021.2002450. 744
672 REFERENCES [23] W. Feng, Y.-G. He, H.-M. Li, and C.-L. Li, ‘‘A plain-image-related chaotic 745

673 [1] Z. Liang, Q. Qin, C. Zhou, N. Wang, Y. Xu, and W. Zhou, ‘‘Medical image image encryption algorithm based on DNA sequence operation and dis- 746

674 encryption algorithm based on a new five-dimensional three-leaf chaotic crete logarithm,’’ IEEE Access, vol. 7, pp. 181589–181609, 2019, doi: 747

675 system and genetic operation,’’ PLoS One, vol. 16, no. 11, Nov. 2021, 10.1109/ACCESS.2019.2959137. 748

676 Art. no. e0260014, doi: 10.1371/journal.pone.0260014. [24] S. Zhou, Z. Zhao, and X. Wang, ‘‘Novel chaotic colour image cryptosys- 749

677 [2] N. Khalil, A. Sarhan, and M. A. M. Alshewimy, ‘‘An efficient tem with deep learning,’’ Chaos, Solitons Fractals, vol. 161, Aug. 2022, 750

678 color/grayscale image encryption scheme based on hybrid chaotic Art. no. 112380, doi: 10.1016/j.chaos.2022.112380. 751

679 maps,’’ Opt. Laser Technol., vol. 143, Nov. 2021, Art. no. 107326, doi: [25] I. Koyuncu, M. Tuna, I. Pehlivan, C. B. Fidan, and M. Alçin, 752
680 10.1016/j.optlastec.2021.107326. ‘‘Design, FPGA implementation and statistical analysis of chaos-ring 753
681 [3] X. Wang, Y. Su, M. Xu, H. Zhang, and Y. Zhang, ‘‘A new image encryption based dual entropy core true random number generator,’’ Analog 754
682 algorithm based on Latin square matrix,’’ Nonlinear Dyn., vol. 107, no. 1, Integr. Circuits Signal Process., vol. 102, no. 2, pp. 445–456, 2020, 755
683 pp. 1277–1293, 2022, doi: 10.1007/s11071-021-07017-7. doi: 10.1007/s10470-019-01568-x. 756

684 [4] N. Munir, M. Khan, A. Al Karim Haj Ismail, and I. Hussain, ‘‘Crypt- [26] A. V. Tutueva, T. I. Karimov, L. Moysis, E. G. Nepomuceno, C. Volos, 757
685 analysis and improvement of novel image encryption technique using and D. N. Butusov, ‘‘Improving chaos-based pseudo-random generators in 758
686 hybrid method of discrete dynamical chaotic maps and Brownian motion,’’ finite-precision arithmetic,’’ Nonlinear Dyn., vol. 104, no. 1, pp. 727–737, 759
687 Multimedia Tools Appl., vol. 81, no. 5, pp. 6571–6584, 2022, doi: 2021, doi: 10.1007/s11071-021-06246-0. 760
688 10.1007/s11042-021-11810-2. [27] T. L. Liao, P. Y. Wan, and J.-J. Yan, ‘‘Design and synchroniza- 761
689 [5] J. M. K. Mastan and R. Pandian, ‘‘Cryptanalysis of two similar chaos-based tion of chaos-based true random number generators and its FPGA 762
690 image encryption schemes,’’ Cryptologia, vol. 45, no. 6, pp. 541–552, implementation,’’ IEEE Access, vol. 10, pp. 8279–8286, 2022, doi: 763
691 Nov. 2021, doi: 10.1080/01611194.2020.1814447. 10.1109/ACCESS.2022.3142536. 764
692 [6] M. Shen, Y. Deng, L. Zhu, X. Du, and N. Guizani, ‘‘Privacy- [28] S. Zhou, X. Wang, W. Zhou, and C. Zhang, ‘‘Recognition of the scale-free 765
693 preserving image retrieval for medical IoT systems: A blockchain-based interval for calculating the correlation dimension using machine learning 766
694 approach,’’ IEEE Netw., vol. 33, no. 5, pp. 27–33, Sep. 2019, doi: from chaotic time series,’’ Phys. A, Stat. Mech. Appl., vol. 588, Feb. 2022, 767
695 10.1109/MNET.001.1800503. Art. no. 126563, doi: 10.1016/j.physa.2021.126563. 768
696 [7] Z. Kuang, Z. Guo, J. Fang, J. Yu, N. Babaguchi, and J. Fan, ‘‘Unnoticeable [29] R. Matthews, ‘‘On the derivation of a, ‘chaotic’ encryption 769
697 synthetic face replacement for image privacy protection,’’ Neurocomput- algorithm,’’ Cryptologia, vol. 13, no. 1, pp. 29–42, 1989, 770
698 ing, vol. 457, pp. 322–333, 2021, doi: 10.1016/j.neucom.2021.06.061. doi: 10.1080/0161-118991863745. 771
699 [8] X. Chai, Y. Wang, Z. Gan, X. Chen, and Y. Zhang, ‘‘Preserving [30] L. M. Pecora and T. L. Carroll, ‘‘Synchronization in chaotic sys- 772
700 privacy while revealing thumbnail for content-based encrypted image tems,’’ Phys. Rev. Lett., vol. 64, no. 8, pp. 821–824, 1990, doi: 10.1063/ 773
701 retrieval in the cloud,’’ Inf. Sci., vol. 604, pp. 115–141, Aug. 2022, doi: 1.4917383. 774
702 10.1016/j.ins.2022.05.008.
[31] J. Fridrich, ‘‘Symmetric ciphers based on two-dimensional chaotic maps,’’ 775
703 [9] Data Encryption Standard (DES), FIPS Standard 46-3, 1999. Int. J. Bifurcation Chaos, vol. 8, no. 6, pp. 1259–1284, 1998, doi: 776
704 [10] Advanced Encryption Standard (AES), FIPS Standard, 197, 2001. 10.1142/S021812749800098X. 777
705 [11] S. H. Zhou, Q. Zhang, X. P. Wei, and C. J. Zhou, ‘‘A summarization
[32] W. Feng, Y.-G. He, H.-M. Li, and C.-L. Li, ‘‘Cryptanalysis of the inte- 778
706 on image encryption,’’ IETE Tech. Rev., vol. 27, no. 6, pp. 503–510,
grated chaotic systems based image encryption algorithm,’’ Optik, vol. 186, 779
707 Nov./Dec. 2010, doi: 10.4103/0256-4602.72583.
pp. 449–457, Jun. 2019, doi: 10.1016/j.ijleo.2018.12.103. 780
708 [12] B. Zhang, D. Xiao, and Y. Xiang, ‘‘Robust coding of encrypted
[33] W. Feng and J. Zhang, ‘‘Cryptanalzing a novel hyper-chaotic image 781
709 images via 2D compressed sensing,’’ IEEE Trans. Multimedia, vol. 23,
encryption scheme based on pixel-level filtering and DNA-level 782
710 pp. 2656–2671, 2021, doi: 10.1109/TMM.2020.3014489.
diffusion,’’ IEEE Access, vol. 8, pp. 209471–209482, 2020, doi: 783
711 [13] X. Wang and Y. Su, ‘‘Image encryption based on compressed sensing and
10.1109/ACCESS.2020.3038006. 784
712 DNA encoding,’’ Signal Process., Image Commun., vol. 95, Jul. 2021,
713 Art. no. 116246, doi: 10.1016/j.image.2021.116246. [34] W. Feng, Z. Qin, J. Zhang, and M. Ahmad, ‘‘Cryptanalysis and improve- 785
ment of the image encryption scheme based on feistel network and dynamic 786
714 [14] C. Zou, X. Wang, and H. Li, ‘‘Image encryption algorithm with matrix
DNA encoding,’’ IEEE Access, vol. 9, pp. 145459–145470, 2021, doi: 787
715 semi-tensor product,’’ Nonlinear Dyn., vol. 105, no. 1, pp. 859–876, 2021,
10.1109/ACCESS.2021.3123571. 788
716 doi: 10.1007/s11071-021-06542-9.
717 [15] C. Wang and Y. Zhang, ‘‘A novel image encryption algorithm with deep [35] S. Liu, C. Guo, and J. T. Sheridan, ‘‘A review of optical image encryption 789

718 neural network,’’ Signal Process., vol. 196, Jul. 2022, Art. no. 108536, doi: techniques,’’ Opt. Laser Technol., vol. 57, pp. 327–342, Apr. 2014, doi: 790

719 10.1016/j.sigpro.2022.108536. 10.1016/j.optlastec.2013.05.023. 791

720 [16] X. Zhang and X. Yan, ‘‘Adaptive chaotic image encryption algorithm based [36] C. L. Li, Y. Zhou, H. M. Li, W. Feng, and J. R. Du, ‘‘Image encryp- 792

721 on RNA and pixel depth,’’ Electronics, vol. 10, no. 15, p. 1770, Jul. 2021, tion scheme with bit-level scrambling and multiplication diffusion,’’ 793

722 doi: 10.3390/electronics10151770. Multimedia Tools Appl., vol. 80, no. 12, pp. 18479–18501, 2021, doi: 794

723 [17] X. Zhang and Z. Gong, ‘‘Color image encryption algorithm based on 3D 10.1007/s11042-021-10631-7. 795

724 zigzag transformation and view planes,’’ Multimedia Tools Appl., vol. 81, [37] G. Chen, Y. Mao, and C. K. Chui, ‘‘A symmetric image encryption 796

725 pp. 31753–31785, Apr. 2022, doi: 10.1007/s11042-022-13003-x. scheme based on 3D chaotic cat maps,’’ Chaos, Solitons Fractals, vol. 21, 797

726 [18] N. Rani, S. R. Sharma, and V. Mishra, ‘‘Grayscale and colored image pp. 749–761, Jul. 2004, doi: 10.1016/j.chaos.2003.12.022. 798

727 encryption model using a novel fused magic cube,’’ Nonlinear Dyn., [38] Y. Mao, G. Chen, and S. Lian, ‘‘A novel fast image encryption scheme 799

728 vol. 108, no. 2, pp. 1773–1796, 2022, doi: 10.1007/s11071-022-07276-y. based on 3D chaotic baker maps,’’ Int. J. Bifurcation Chaos, vol. 14, no. 10, 800

729 [19] Y. Naseer, T. Shah, and D. Shah, ‘‘A novel hybrid permutation substitu- pp. 3613–3624, Oct. 2004, doi: 10.1142/S021812740401151X. 801

730 tion base colored image encryption scheme for multimedia data,’’ J. Inf. [39] J. Chen, L. Yu Zhang, and Y. Zhou, ‘‘Re-evaluation of the security 802
731 Secur. Appl., vol. 59, Jul. 2021, Art. no. 102829, doi: 10.1016/j.jisa.2021. of a family of image diffusion mechanisms,’’ IEEE Trans. Circuits 803
732 102829. Syst. Video Technol., vol. 31, no. 12, pp. 4747–4758, Dec. 2021, doi: 804

733 [20] L. Qu, F. Chen, S. Zhang, and H. He, ‘‘Cryptanalysis of reversible 10.1109/TCSVT.2021.3054508. 805

734 data hiding in encrypted images by block permutation and co- [40] L. Liu, Y. Lei, and D. Wang, ‘‘A fast chaotic image encryption scheme 806
735 modulation,’’ IEEE Trans. Multimedia, vol. 24, pp. 2927–2937, 2021, doi: with simultaneous permutation-diffusion operation,’’ IEEE Access, vol. 8, 807
736 10.1109/TMM.2021.3090588. pp. 27361–27374, 2020, doi: 10.1109/ACCESS.2020.2971759. 808

96004 VOLUME 10, 2022

 B. Ge et al.: Fast Chaotic Image Encryption Algorithm Using a Novel Divide and Conquer Diffusion Strategy

809 [41] T. Li, J. Shi, and D. Zhang, ‘‘Color image encryption based on joint per- [54] Y. Wu, Y. Zhou, G. Saveriades, S. Agaian, J. P. Noonan, and 854
810 mutation and diffusion,’’ Proc. SPIE, vol. 30, no. 1, 2021, Art. no. 013008, P. Natarajan, ‘‘Local Shannon entropy measure with statistical tests for 855
811 doi: 10.1117/1.JEI.30.1.013008. image randomness,’’ Inf. Sci., vol. 222, pp. 323–342, Feb. 2013, doi: 856
812 [42] C. Fu, J.-J. Chen, H. Zou, W.-H. Meng, Y.-F. Zhan, and Y.-W. Yu, 10.1016/j.ins.2012.07.049. 857
813 ‘‘A chaos-based digital image encryption scheme with an improved dif- [55] Z. Li, C. Peng, L. Li, and X. Zhu, ‘‘A novel plaintext-related image 858
814 fusion strategy,’’ Opt. Exp., vol. 20, no. 3, pp. 2363–2378, 2012, doi: encryption scheme using hyper-chaotic system,’’ Nonlinear Dyn., vol. 94, 859
815 10.1364/OE.20.002363. no. 2, pp. 1319–1333, 2018, doi: 10.1007/s11071-018-4426-4. 860
816 [43] X. Zhang and Z. Zhao, ‘‘Chaos-based image encryption with total shuf-
817 fling and bidirectional diffusion,’’ Nonlinear Dyn., vol. 75, nos. 1–2,
818 pp. 319–330, Jan. 2014, doi: 10.1007/s11071-013-1068-4.
819 [44] H. X. Zhao, S. C. Xie, J. Z. Zhang, and T. Wu, ‘‘Efficient image encryption BIN GE received the Ph.D. degree in engineer- 861

820 using two-dimensional enhanced hyperchaotic Henon map,’’ Proc. SPIE, ing from the University of Chinese Academy of 862

821 vol. 29, no. 2, 2020, Art. no. 023007, doi: 10.1117/1.JEI.29.2.023007. Sciences, in 2016. He is currently a Lecturer 863
822 [45] X. Wang and H. Zhao, ‘‘Fast image encryption algorithm based on paral- with Nantong Vocational University. His current 864
823 lel permutation-and-diffusion strategy,’’ Multimedia Tools Appl., vol. 79, research interests include the theory and applica- 865
824 no. 27, pp. 19005–19024, 2020, doi: 10.1007/s11042-020-08810-z. tion of cryptography and chaotic cryptography. 866
825 [46] W. Feng, X. Zhao, J. Zhang, Z. Qin, J. Zhang, and Y. He, ‘‘Image
826 encryption algorithm based on plane-level image filtering and discrete
827 logarithmic transform,’’ Mathematics, vol. 10, no. 15, p. 2751, 2022, doi:
828 10.3390/math10152751.
829 [47] K. Qian, W. Feng, Z. Qin, J. Zhang, X. Luo, and Z. Zhu, ‘‘A novel
830 image encryption scheme based on memristive chaotic system and
831 combining bidirectional bit-level cyclic shift and dynamic DNA- ZHIHUA SHEN received the Ph.D. degree in engi- 867
832 level diffusion,’’ Frontiers Phys., vol. 10, p. 718, Aug. 2022, doi: neering from Xi’an Jiao Tong University, in 2017. 868
833 10.3389/fphy.2022.963795. He is currently a Lecturer with Nantong Vocational 869
834 [48] S. Zhou, X. Wang, Y. Zhang, B. Ge, M. Wang, and S. Gao, ‘‘A novel University. His current research interests include 870
835 image encryption cryptosystem based on true random numbers and chaos theory and circuit design. 871
836 chaotic systems,’’ Multimedia Syst., vol. 28, pp. 95–112, May 2021, doi:
837 10.1007/s00530-021-00803-8.
838 [49] L. Teng, X. Wang, F. Yang, and Y. Xian, ‘‘Color image encryption based
839 on cross 2D hyperchaotic map using combined cycle shift scrambling and
840 selecting diffusion,’’ Nonlinear Dyn., vol. 105, no. 2, pp. 1859–1876, 2021,
841 doi: 10.1016/j.optlastec.2021.107252.
842 [50] X. Gao, ‘‘Image encryption algorithm based on 2D hyperchaotic map,’’
843 Optics Laser Technol., vol. 142, Oct. 2021, Art. no. 107252, doi:
JINBAO ZHANG received the Ph.D. degree in 872
844 10.1016/j.optlastec.2021.107252.
communication and information system from the 873
845 [51] F. Yuan, Y. Li, and G. Wang, ‘‘A universal method of chaos cascade and its
846 applications,’’ Chaos, Interdiscipl. J. Nonlinear Sci., vol. 31, no. 2, 2021, Nanjing University of Aeronautics and Astronau- 874

847 Art. no. 021102, doi: 10.1063/5.0041518. tics, in 2019. He is currently a Lecturer with 875

848 [52] M. Haahr. RANDOM.ORG: True Random Number Services. Nantong University. His current research interests 876

849 Accessed: Jun. 28, 2022. [Online]. Available: https://www.random.org include physical security issues, side-channel anal- 877

850 [53] F. Pareschi, R. Rovatti, and G. Setti, ‘‘On statistical tests for randomness ysis for cryptographic hardware and embedded 878

851 included in the NIST SP800–22 test suite and based on the binomial dis- systems, and application-specified integrated cir- 879
852 tribution,’’ IEEE Trans. Inf. Forensics Security, vol. 7, no. 2, pp. 491–505, cuit design. 880
853 Apr. 2012, doi: 10.1109/TIFS.2012.2185227. 881

VOLUME 10, 2022 96005