1. Introduction
Nowadays, individuals with various profiles often participate in the sharing of multimedia material such as images on the internet and other communication networks. To ensure confidential transmission and the accuracy of the received data, techniques designed to secure this type of information play a crucial role. Image encryption differs from text encryption owing to inherent properties such as the high correlation between neighbouring pixels and redundancy. Thus, the design of robust secure, resilient, and effective image encryption algorithms will always remain at the forefront of commercial and scientific endeavours.
Many image encryption techniques have been suggested and developed by researchers, such as Arnold transformation-based [
1,
2,
3], quantum system-based [
4,
5,
6], compression sensing-based [
7,
8,
9], cellular automata-based [
10,
11,
12], DNA computing-based [
13,
14,
15], elliptic curves-based [
16,
17,
18], and chaotic system-based [
19,
20,
21,
22,
23] approaches. The approach of digital image encryption based on chaotic systems is considered to be more effective than other approaches due to their ability to improve unpredictability, which makes encryption more resistant to attacks. These systems are generally categorised into one-dimensional systems such as Logistic map [
24], higher-dimensional chaotic systems such as Lorenz system [
25] and Rössler system [
26], and fractional-order systems such as Chen system [
27] (for more details, see [
28]). The higher-dimensional systems outperform low-dimensional ones in terms of randomness, nonlinearity, and generated key sequences, which contain a large keyspace.
In the literature, the use of multidimensional chaotic systems for image encryption is well documented. For instance, Naseer et al. [
29] introduced a method for image encryption that uses a three-dimensional chaotic system and a permutation box for the substitution phase and an S-box for the permutation phase. However, the provided image encryption does not achieve plain image sensitivity. Maazouz et al. [
30] introduced a novel three-dimensional chaotic system and explained its function in creating S-boxes, using the resulting S-boxes to develop a new image encryption method. However, the final cyphered image is produced after encryption several rounds. Zheng et al. [
31] introduced a new image encryption that combines the improved two-dimensional logistic sine map and dynamic DNA sequence encryption. They showed the DNA scheme’s security problem, which does not change during the encryption process, can be solved using their approach. Kumar and Girdhar [
32] suggested an image encryption technique with a confusion stage depending on two-dimensional Logistic maps and a diffusion stage depending on a Lorenz–Rossler chaotic system. Wang et al. [
33] employed a six-dimensional hyperchaotic system in conjunction with DNA coding to propose an image encryption method, which executes diffusion encryption in multiple rounds. Hu et al. in [
34] suggested a Fibonacci dynamical system-based cloud model-based colour image encryption technique combined with a matrix convolution algorithm. This combination improves the security of encrypted images in terms of the randomness of bits.
An assessment of the literature reveals that several colour image encryption methods have security vulnerabilities, such as susceptibility to noise attacks, low keyspace, and failure to withstand statistical attacks. In addition, encrypting colour images efficiently is a significant challenge, since they have more data than greyscale images. These limitations motivated us to develop a three-dimensional chaotic system with complex dynamic behaviour to ensure desired properties, such as dissipative and great sensitivity to small changes in control parameters and initial conditions. Such properties are analogous to image encryption requirements. The main aim was to design a novel encryption method utilising the proposed system and the self-invertible matrix generated by a modified Diffie–Hellman key exchange protocol.
This paper’s contributions are outlined below:
The proposed conservative system is employed, where its initial conditions depend on the input image to create the first set of secret keys required to scramble the plain image;
The first use of the Diffie–Hellman key exchange protocol with a self-invertible matrix;
The modified Diffie–Hellman key exchange protocol is proposed and combined with the self-invertible matrix of size to obtain the second set of keys;
Based on the shared secret key matrix , the diffusion stage is performed;
One of the main drawbacks of the matrix generated by the modified Diffie–Hellman key exchange protocol is that the inverse of the matrix does not always exist, which affects the decryption process. To avoid this problem, we utilised the self-invertible matrix, which reduces the computational process requirements during decryption and increases the unpredictability of pixels distribution during encryption;
The keyspace size is increased to by using two sets of keys, making brute-force assaults impossible;
The three-dimensional conservative chaotic system, together with the key matrix , provides a high level of security;
The suggested method was evaluated against several assaults to measure how well it performs in terms of robustness, effectiveness, and cryptanalysis resistance.
This paper is structured as follows. The stability of the proposed conservative system is investigated and possible types of equilibria are determined in
Section 2. The conservativeness of the proposed system is also discussed in this section. In
Section 3, a modified Diffie–Hellman Key Exchange Protocol, which was combined with the proposed conservative system, is provided to design a new image encryption technique. In
Section 4, the proposed encryption method is described. The security evaluation and results are presented in
Section 5. A summary of the entire study is presented in
Section 6 as its conclusion.
2. Conservative Chaotic System
Chaotic systems can be found in different areas such as weather forecasting [
35], telecommunications [
36], and biological modelling [
37], and can be classified into different categories by physical (number of dimensions), dynamical (number of wings), and algebraic features (number of equilibrium points).
It is well known that almost all classical chaotic systems are dissipative, while a few are conservative systems. The conservative ones are an old category of dynamical systems studied by Euler, Lagrange, and Hamilton, whose phase space volume is conserved. The characteristics of conservative systems are as follows:
The sum of finite-time local Lyapunov exponents is zero;
Divergence of the vector field is zero;
Local Lyapunov dimension is equal to the order of the system;
The Hamiltonian energy of the system is invariable.
Sprott in [
38] raised a fascinating question of whether new chaotic dynamical systems are still needed. His answer to the question motivated us to propose a novel chaotic system with some properties and applications as follows:
where
are the system parameters and
are system variables.
There are several chaotic systems based on the system in [
38] with different features, but our proposed chaotic system has the following unique characteristics:
The system is conservative for but dissipates for and . Note that the semianalytical and seminumerical technique was used to demonstrate the conservativeness of the new system;
The proposed system (
1) has a stable or unstable line, or no equilibria;
Coexisting attractors appear.
The properties of the proposed conservative system to analyse the global dynamics are explored in the following two sections.
2.1. Dynamical Analysis
Let us consider the following algebraic equations:
Observe that system (
1) has a line of equilibria at
if
, or else it has no equilibrium for either
or
. In the case of no equilibrium, the proposed system can be classified as a chaotic system with a hidden strange attractor, a basin of attraction that does not contain neighbourhoods of equilibria.
Additionally, to set the stability of system (
1), we consider the Jacobian matrices at the possible equilibrium point, i.e., at the origin
when
. The Jacobian matrix of the proposed system at the origin point is given by
and the corresponding characteristic polynomial of
is
where
, while the eigenvalues of
are
and
According to the Routh–Hurwitz stability criterion, it is easy to see that system (
1) has stable equilibrium. Obviously, in the case of no equilibrium, the proposed system (
1) shows a symmetry around the
-axis, such that
. The system’s attractors are depicted in
Figure 1 with initial data
and
and
When
and
the conservativeness of system (
1) can be testified by the divergence as follows:
By (
5), the conservativeness is not clear, since the dissipation is shown by the time-averaged value
through the trajectory. It was established in [
39] that the average value (
) of any function can be computed by
One can verify that the average of
of system (
1) is zero. Furthermore, the Lyapunov exponents (LEs) of system (
1) are
, with
(see
Figure 2). The Kaplan–Yorke fractional dimension presents the complexity of the attractor and is defined by
where
j is the largest integer satisfying
and
. The corresponding fractional dimension of system (
1) is 3. Therefore, this indicates a strange attractor. Thus, the dimension is equivalent to the number of system variables, showing the proposed system’s conservativeness. System (
1) is conservative, with dynamics shown in
Figure 1.
For
and
system (
1) is invariant to the transformation
; hence, it is time-reversible with LEs that are symmetric around zero. Additionally, a positive definite energy function in quadratic form can be provided as the Hamiltonian function:
Considering the time derivative of
we have
which demonstrates that, if
is zero, the energy of system (
1) is invariable. The system is, thus, conservative by nature. It should be noted that any attempt to simplify system (
1) will result in a non-conservative system, as illustrated in
Table 1.
The bifurcation diagram was obtained by plotting the local maxima of the state variable
when
and changing the value of
in the interval
, and, also, when
and changing the value of
in the interval
, as shown in
Figure 2.
Table 1.
Two modified chaotic systems of system (
1).
Table 1.
Two modified chaotic systems of system (
1).
Modified System | LEs | | Phase Portrait |
---|
| | 2.486 | Figure 3a |
|
|
| | 2.498 | Figure 3b |
|
|
2.2. Coexisting Attractors
To investigate the phenomena of coexisting attractors, we need to check the influence of the parameters and initial conditions on the behaviours of system (
1). We have observed that, if
or
, the proposed system reveals chaotic dynamics with no equilibrium.
Figure 4 displays a few coexisting attractors of system (
1).
Figure 4a shows a chaotic attractor coexists, whereas
Figure 4b shows quasi-periodic coexistence.
Note that, according to the eigenvalues obtained at the origin, we have Hopf-zero equilibrium, but there is no Hopf-zero bifurcation at the origin coordinates (i.e., there is no periodic solution (attractor) that may coexist except the trivial solution at the origin); this fact can be easily proven by the average method.
Figure 1.
Phase portrait of the proposed system (1) when and .
Figure 1.
Phase portrait of the proposed system (1) when and .
Figure 2.
(
a) Bifurcation diagram of the novel system (
1) when varying parameter
. (
b) Bifurcation diagram of the novel system (
1) when varying parameter
. (
c) Lyapunov exponents of the system (
1) when
and
.
Figure 2.
(
a) Bifurcation diagram of the novel system (
1) when varying parameter
. (
b) Bifurcation diagram of the novel system (
1) when varying parameter
. (
c) Lyapunov exponents of the system (
1) when
and
.
Figure 3.
Two modified systems of system (
1).
Figure 3.
Two modified systems of system (
1).
Figure 4.
(a) A chaotic attractor coexists when and (b) Quasi-periodic coexist when and .
Figure 4.
(a) A chaotic attractor coexists when and (b) Quasi-periodic coexist when and .
3. Modified Diffie–Hellman Key Exchange Protocol (MDHKEP)
Diffie–Hellman key exchange agreement was established by Diffie and Hellman in the 1970s [
40]. A modification of the protocol based on the original one was introduced as follows. Assume two users
A and
B decide to communicate over an insecure channel. They should first choose the two public parameters shown below:
p—a large prime number;
G = —a random matrix, such that and .
Thereafter, User
A randomly picks his private key
from the interval
and calculates his public key as follows:
User
B also randomly picks his private key
and calculates his public key as follows:
By doing so, users
A and
B will be able to successfully exchange the parameters required to compute the initial shared key, as shown below:
Next, users
A and
B create the final shared secret key matrix
by employing the suggested technique in Acharya et al. [
41], where each user can build his
self-invertible key matrix
. Let
be a self-invertible matrix that can be partitioned as
such that
The value of each entry (i.e., , , and ) can be found by solving the following equations: , , and , where is the identity matrix.
5. Experiments and Performance Analysis
The trial simulation setup consisted of a computer with a Windows 11 operating system, 8 GB of RAM, a 1.30 GHz central processing unit, and Matlab R2018A. In these experiments, we used Airplane, Baboon, Lena, Peppers, and Tree images with sizes of and as the plaintext images.
5.1. Encryption and Decryption Results
The encryption and decryption results for the proposed method are displayed in
Figure 7. As shown, the decrypted images were similar to the original image, whilst the cypher images were noise-like and unidentifiable, which shows that the proposed approach has effective encryption and decryption outcomes. We next checked the image encryption scheme against some possible assaults.
5.2. Statistical Analysis
In this section, we show results from checking the following statistical indices:
- I.
Histogram and Chi-Square Test
To visually describe the distribution of pixel intensities in the image, we need to consider the image histogram that gives a graphic view. The fact that pixel intensities are constrained to a given range causes histograms for original images to often be inconsistent. Thus, cryptanalysts can leverage this characteristic to intercept the encryption by utilising histogram-based attacks.
Figure 8 depicts the colour image, Peppers, and histograms of red, green, and blue channels that indicate the encrypted analogue of the original image. The original image has non-uniform histograms, but the encrypted image has highly uniform distributions of pixel intensities in all channels, which shows the suggested method can withstand statistical attack. The Chi-square test was employed to determine the homogeneity of a histogram and is calculated via the following equation:
where
represents the grey recurrence value of
i and
The value of
was 293.2478, with
as the significance level and
as the degree of freedom [
22].
Table 2 shows the
values of the cyphered image. Since all the values were less than 293, the histogram of the images encrypted with our suggested approach was uniform.
- II.
Entropy
One of the most significant measures in dynamical systems theory and security is entropy, which is a statistical test for calculating randomness in an image, defined as
where
is the probability of the element
, and
T represents the total number of elements
. For an encrypted image, eight is the ideal value of entropy [
16]. In this paper, we used the suggested approach to encrypt colour images and determine the entropy values of the cyphered images, as shown in
Table 3. In a second experiment, we used our technique and some other existing methods [
2,
9,
12,
18,
23] to encrypt Lena’s colour image.
Table 4 displays the computed entropy values. In general, the average entropy for the cyphered image produced by our approach beat the outcomes of the other techniques [
2,
9,
12,
23] or was similar to the outcomes of the other techniques [
18]. The results indicate that the proposed method has higher unpredictability of image information compared to other existing methods, as
Table 4 shows.
5.3. Keyspace Analysis
For a good encryption scheme, the keyspace must be large enough to withstand brute-force assaults. The suggested algorithm’s key consists of the starting values
, and
, as well as the parameters
and
. The keyspace size can reach
if the accuracy is
. The algorithm also uses the keys
, and
. The size of these keys depends on the value of
p in
, where the keys
are of size
, whereas those for
and
are of size
. The shared secret key
has a size of 128 bits, since it is a
matrix whose elements of 8 bits belong to the field
. The keyspace analysis for the suggested algorithm may, thus, be determined as the following:
. Attackers who select a 16-bit value for
p will need to carry out
operations to defeat the method, which is robust enough to withstand brute-force attacks.
Table 5 compares the obtained keyspace of the suggested image encryption technique with the literature’s techniques. The findings in
Table 5 demonstrate that the suggested technique has a larger keyspace than existing methods [
2,
9,
12,
18,
23].
5.4. Key Sensitively Test
A robust image encryption scheme must be sensitive to even minor changes in the initial conditions of the secret key employed. The recovered image becomes noisy and scrambled when this key is slightly changed. The private key parameters were (
,
,
,
,
,
,
). The plain image of Peppers was encrypted in this experiment using the private key, and the result is displayed in
Figure 9(a1). The original image was not recovered as in
Figure 9(a2) when the cyphered image was decrypted in
Figure 9(a1) by altering
(one of the private keys) to
while keeping the other parameters the same as above. The original image was only recovered with the correct private key, as shown in
Figure 9(a3).
5.5. Anti-Differential Attack
The ability to withstand differential attack is a crucial measure of plaintext sensitivity. Attackers typically alter a plain image’s pixels slightly and attempt to determine differences between the original encrypted image and the modified encrypted image to find a weakness in the encryption algorithms, such as attaining information about the secret key. Two measures, denoted by NPCR and UACI, can be used to test the ability to resist differential attacks. The equations of NPCR and UACI are as follows:
and
where
and
are the cypher images created by two original images with a one-pixel difference, and
is mathematically expressed by
To guarantee the security of the algorithm, the UACI value must be near
, and NPCR must be near
.
Table 6 and
Table 7 display the comparative results of UACI and NPCR with respect to other methods [
2,
9,
12,
18,
23]. The average values of UACI and NPCR for the cyphered image Lena outperformed the techniques [
9,
12,
18,
23] in terms of UACI and outperformed the techniques [
2,
9,
12,
23] in terms of NPCR. The results reveal that UACI and NPCR average values were perfect, so that the proposed approach can effectively survive differential attacks.
5.6. Noise Attack
When images are transmitted over media channels, they are often subjected to noise. If the cyphered image is exposed to noise, it may influence the image quality after decryption. We contaminated the cypher image “Peppers” of size
with salt and pepper noise with densities
and
; the analogous decrypted images are shown in
Figure 10a–a2. From the below figures, one can see that these images were noisy but readable.
5.7. Correlation Analysis
Let
represent the neighbouring pixels. In the usual images,
values are very close to each other, which means neighbouring pixels are highly correlated in the original images. This characteristic may be exploited by attackers to break the cypher scheme. Thus,
have to be highly uncorrelated in the encrypted images. The horizontal, vertical, and diagonal correlation coefficient between any two adjacent pixels,
x and
y, is calculated by
where
indicates the expected values of the random variables,
and
represent greyscale values of
, and
denotes the total number of pixels involved in the computations.
In
Figure 11, one can observe that the coefficients of correlation in the cypher images in all three (horizontal, vertical, and diagonal) directions were almost zero. Therefore, the connected pixels in the cyphered images were highly uncorrelated. The findings displayed in
Table 8 show that the correlation values of the various colour images were all close to 0, indicating that the proposed method cracked the strong correlation between neighbouring pixels in all tested images.
Table 9 shows a comparison of correlation values of the suggested encryption method and existing encryption methods. One can see that our technique’s correlation values were closer to zero, indicating that the proposed method has superior protection against statistical attacks.
Table 8.
Correlation coefficient test results of different encrypted images in three channels.
Table 8.
Correlation coefficient test results of different encrypted images in three channels.
Image | Size | Directions | R | G | B |
---|
Airplane | | H | 0.0032 | −0.0011 | 0.0010 |
V | 0.0035 | 0.0006 | −0.0008 |
D | −0.0002 | 0.0033 | 0.0018 |
Baboon | | H | −0.0018 | −0.00007 | 0.0020 |
V | −0.0004 | −0.00100 | −0.0008 |
D | 0.0016 | −0.00030 | 0.0031 |
Lena | | H | −0.00006 | −0.0003 | −0.0018 |
V | 0.00150 | −0.0020 | −0.0011 |
D | 0.00010 | −0.0011 | 0.0021 |
Peppers | | H | −0.0004 | 0.0016 | 0.0025 |
V | 0.0006 | 0.0032 | 0.0016 |
D | −0.0022 | 0.0006 | 0.0043 |
Tree | | H | 0.0002 | −0.0010 | −0.0018 |
V | 0.0014 | 0.0025 | 0.0008 |
D | −0.0006 | −0.0002 | −0.0029 |
Airplane | | H | −0.0067 | −0.0024 | 0.0062 |
V | 0.0057 | −0.0047 | −0.0007 |
D | −0.0004 | −0.0079 | −0.0002 |
Baboon | | H | 0.0020 | −0.0072 | 0.0043 |
V | −0.0037 | −0.0044 | 0.0039 |
D | −0.0043 | 0.0062 | 0.0012 |
Lena | | H | −0.00004 | 0.0004 | −0.0027 |
V | −0.00370 | −0.0044 | 0.0039 |
D | 0.00400 | 0.0060 | −0.0002 |
Peppers | | H | −0.0014 | −0.0072 | 0.0048 |
V | 0.0022 | 0.0070 | −0.0010 |
D | 0.0020 | 0.0062 | −0.0039 |
Tree | | H | 0.0051 | −0.0017 | −0.0002 |
V | −0.0055 | 0.0085 | −0.0014 |
D | 0.0053 | 0.0021 | −0.0067 |
Table 9.
Comparison of correlation coefficients of encrypted Lena image in three channels.
Table 9.
Comparison of correlation coefficients of encrypted Lena image in three channels.
Scheme | Size | Directions | R | G | B |
---|
Proposed | | H | −0.00004 | 0.0004 | −0.0027 |
V | −0.00370 | −0.0044 | 0.0039 |
D | 0.00400 | 0.0060 | −0.0002 |
[2] | | H | 0.00040 | −0.0050 | −0.0103 |
V | −0.00140 | −0.0144 | 0.0010 |
D | 0.00200 | 0.0073 | −0.0040 |
[9] | | H | −0.00200 | −0.0013 | −0.0059 |
V | −0.00220 | −0.0041 | 0.0014 |
D | 0.00690 | 0.0059 | 0.0035 |
[12] | | H | 0.00080 | 0.0002 | 0.0274 |
V | 0.00130 | −0.0001 | −0.0002 |
D | 0.00060 | −0.0048 | −0.0055 |
[18] | | H | 0.00090 | 0.0007 | 0.0007 |
V | −0.00090 | −0.0009 | −0.0007 |
V | 0.00170 | 0.0009 | 0.0009 |
[23] | | H | 0.00073 | 0.0031 | −0.0051 |
V | −0.00054 | 0.0008 | 0.0033 |
D | 0.00147 | −0.0015 | −0.0062 |