2014 International Conference on Control, Instrumentation, Energy & Communication(CIEC) 583

Colour Image Encryption Based on Multiple

Fractional Order Chaotic Systems
Tanmoy Dasgupta∗ , Pritam Paral† , and Samar Bhattacharya‡
Dept. of Electrical Engineering, Jadavpur University, Kolkata, India
∗ [email protected], † [email protected], ‡ samar [email protected]

Abstract—This paper demonstrates a colour image encryption A. Fractional-Order Lorenz’s System

algorithm that incorporates multiple fractional order chaotic
systems. Here, the fractional order extensions of four well known The fractional order extension of Lorenz’s system is given
chaotic systems are tactfully combined together to generate the by (2)
cyphering key. The encrypted image shows uniform histogram,
zero autocorrelation and very high entropy for all the three Dq1a x1a = σ (x2a − x1a ), (2a)
colour-channels. Moreover, the algorithm possesses an extremely Dq2a x2a = x1a (ρ − x3a ) − x2a , (2b)
huge key-space. Thus, the algorithm becomes invulnerable against Dq3a x3a = x1a x2a − β x3a . (2c)
brute-force attacks. However, due to the fact that the fractional
order derivatives are not ‘local operations’ like their integer
order counterparts, solving fractional order differential equations Consider the values of the parameters to be σ = 10,
become a difficult and time consuming task for computers. To ρ = 28 and β = 8/3, and the derivative orders to be (the
overcome this, some efficient computing techniques have been same derivative orders will be considered for all the remaining
adopted to run the encryption-decryption process reasonably fast. systems)    
q1a 0.95
qa = q2a = 0.95 , (3)
Keywords—Chaos, fractional order systems, image encryption.
q3a 0.95
along with the following initial conditions.
I. I NTRODUCTION    
x1a (0) 0.1
Use of chaotic systems in image encryption was first xa (0) = x2a (0) = 0.1 . (4)
proposed by Matthews [1]. Since then, many image encryption x3a (0) 0.1
algorithms have been developed based on chaotic systems
[2]–[4] . Since the last few years, fractional order extensions Fig. 1 shows the corresponding state trajectories of the
of the chaotic systems have become quite popular choice Lorenz’s system.
over the integer order versions because they provide much
larger key-spaces [5], [6]. In the present work, we propose B. Fractional-Order Chen’s System
and demonstrate a scheme, in which multiple chaotic systems
are combined to accomplish the algorithm. In this work, the The fractional order Chen’s system is described by (5) as
fractional order versions of four well known three-dimensional
chaotic systems [7], viz. Lorenz’s system, Chen’s system, Liu’s Dq1b x1b = a (x2b − x1b ), (5a)
system and Volta’s system are used. The states of these chaotic Dq2b x2b = (c − a) x1b − x1b x3b + c x2b , (5b)
systems are merged so as to raise the degree of randomness in Dq3b x3b = x1b x2b − b x3b . (5c)
the encryption process and thus increasing the key sensitivity.
First, the above mentioned chaotic systems are described The values of the parameters are supposed to be a = 35,
in brief. Then, the proposed image encryption algorithm have b = 3 and c = 28. Now, the initial conditions are taken as
   
been discussed in detail. x1b (0) −9.0
xb (0) = x2b (0) = −15.0 . (6)
x3b (0) 14.0
II. F RACTIONAL O RDER C HAOTIC S YSTEMS
Throughout this paper, the fractional (q-th) order Fig. 2 shows the corresponding state trajectories of the
Grünwald-Letnikoff (G-L) derivative [8] Dq f (t) is considered, Chen’s system.
which is defined as
t/h   C. Fractional-Order Liu’s System
 q
(−1) r
f (t − rh) The fractional order generalization of the Liu’s system is
r=0
r given as (7)
D f (t) = lim+
q
. (1)
h→0 hq Dq1c x1c = −a x1c − e x22c , (7a)
Dq2c x2c = b x2c − k x1c x3c , (7b)
The fractional order chaotic systems that are used in this
work are briefly discussed as follows. Dq3c x3c = −c x3c + m x1c x2c . (7c)

978-1-4799-2044-0/14/$31.00©2014IEEE

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 12,2022 at 08:34:21 UTC from IEEE Xplore. Restrictions apply.
 2014 International Conference on Control, Instrumentation, Energy & Communication(CIEC) 584

Fig. 1. State trajectories for the Lorenz’s system for 50 s. Fig. 4. State trajectories for the Volta’s system for 20 s.

D. Fractional-Order Volta’s System

The fractional order Volta’s system is defined as in (9)
Dq1d x1d = −x1d − a x2d − x3d x2d , (9a)
Dq2d x2d = −x2d − b x1d − x1d x3d , (9b)
Dq3d x3d = c x3d + x1d x2d + 1. (9c)

The values of the parameters are taken as a = 5, b = 85

and c = 0.5, along with the initial conditions as follows
   
x1d (0) 8.0
xd (0) = x2d (0) = 2.0 . (10)
x3d (0) 1.0
Fig. 2. State trajectories for the Chen’s system for 50 s.
Fig. 4 shows the corresponding state trajectories of the
Volta’s system.
In this work, the Liu’s system is considered with the
parameters a = e = 1, b = 2.5, k = m = 4 and c = 5. III. T HE I MAGE E NCRYPTION –D ECRYPTION A LGORITHM
Also, the initial conditions are set to be In this paper, the proposed image encryption algorithm
    encompasses two phases.
x1c (0) 0.2
xc (0) = x2c (0) = 0.0 . (8) In the first phase, the solutions of the adopted fractional
x3c (0) 0.5 order chaotic systems, given by (2), (5), (7) and (9) are
calculated for a certain amount of time. Thereafter, the state-
time chronicle is manipulated so that an array (of adequate
size) of pseudo-random 8-bit unsigned integers is produced.
Fig. 3 shows the corresponding state trajectories of the
Liu’s system. In the second phase, the image pixels of the original colour
image are manipulated with the help of the pseudo-random
sequence generated in phase-1.
The phases have been discussed in detail.

A. Phase-1: Generation of Pseudo-random Sequence

Step–1: Compute how many pixels are present in the
original image.
Step–2: Fractional order differential equations in (2), (5),
(7) and (9) should be iterated for appropriate number of times.
This is done to ensure that the array generated is of sufficient
size so as to manipulate all the pixels of the original image.
Solver methods which possess the ability to overcome transient
conditions are preferred in this matter.
Step–3: Now, considering the fact that the state chronicle
Fig. 3. State trajectories for the Liu’s system for 50 s. obtained so far will have 64-bit floating point data-type, and on

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 12,2022 at 08:34:21 UTC from IEEE Xplore. Restrictions apply.
 2014 International Conference on Control, Instrumentation, Energy & Communication(CIEC) 585

an average, this data type can hold about 14 to 15 digits after Thus the expressions of XR , XG and XB will get altered
the decimal point, the pseudo-random arrays of 8-bit integers for different values of the parameter that controls the selection
are produced by using (11) of states.
 
Xi = (|xi | − xi ) × 1014 (mod 256). (11) Step–2: Taking the values of Xk (where k = R, G, B) in
order, a key-image [Ki,j,k ]m×n×p is constructed which takes
Where Xi denotes the pseudo-random integer correspond- the principal role in this encryption business. Here an instance
ing to a state value. With the help of above procedure, a has been set up,
very useful set of pseudo-random 8-bit unsigned integers can K0,0,0 = XR (0)
be obtained. These integers vary between 0 and 255 and
follow an almost uniform distribution throughout the range. K0,0,1 = XG (0)
Here, the initial conditions as well as the order of derivatives K0,0,2 = XB (0)
can be selected arbitrarily. As a consequence, a total of six
control parameters (from each chaotic system) is found to Repeat this process as
be available to form the key space. Hence, the key space
becomes as huge as (1014 )4×6 = 10336 . Now, if the control K0,1,0 = XR (h)
parameters are slightly changed, the pseudo-random sequence K0,1,1 = XG (h)
thus generated will be heavily affected just because of the fact K0,1,2 = XB (h)
that the response of fractional order chaotic systems are heavily
dependent on the initial conditions as well as the derivative and so on, continuing until all the elements of the key-image
orders. [Ki,j,k ]m×n×p are loaded. Here, h is the computation step used
to solve (2), (5), (7) and (9).
B. Phase-2: Image Pixel Manipulation Now, by performing bitwise XOR operation between each
Let the original colour image be denoted by [Ti,j,k ]m×n×p original image pixel and the corresponding key image pixel,
which is nothing but an array of 8-bit unsigned integers of the encrypted version of the original colour image which is
order m×n×p, where m = row length, n = column length and indicated by [Ri,j,k ], is produced. Hence, the expression for
p = number of channels of image. In this scheme, the pseudo- [Ri,j,k ] can be written as
random arrays corresponding to the states of four fractional [Ri,j,k ] = [Ti,j,k ] ⊕ [Ki,j,k ]. (12)
order chaotic systems previously discussed are combined, in a
fashion, to generate the cyphering keys for channel-1, channel- C. Image Decryption
2 and channel-3 (henceforth called R, G and B channels
respectively). Here we accomplish the pixel manipulation The decryption procedure follows the principle that
routine through following two steps: (X ⊕ Y ) ⊕ Y = X. (13)
Step–1: At first, we form a table named State Allocation
Table in which the states of the chaotic systems, after convert- Hence, the algorithm can conveniently decrypt the en-
ing them into 8-bit unsigned integers, are distributed under crypted image just by executing bitwise XOR operation be-
different channels. tween the encrypted image and the same key-image.
Here, we apply a trick. Instead of allocating the sates
exactly in the same order every time, some switch-cases are IV. E XPERIMENTAL R ESULTS
created which will jumble up the state selection. The switch- Throughout this work, Python 2.7 (along with NumPy,
cases will be controlled by another control parameter. This Matplotlib and Cython) is used as the experimental platform.
increases the key space further. An instance of such a modified However, the numerical solution of even a very simple one di-
allocation table is presented in Table I. mensional fractional order homogeneous differential equation
like Dq y(t) + b y(t) = 0 becomes as complicated as [9]
TABLE I. A N INSTANCE OF THE S TATE A LLOCATION TABLE
 r  
j q
Chaotic States allocated to different Channels y(rh) = −bhq y(rh − h) − (−1) y(rh − jh), (14)
System R-Channel G-Channel B-Channel
j=1
j
Lorenz X1a X2a X3a
X3b X1b X2b
Chen
Liu X2c X1c X3c
for r = n, n + 1, · · · . So, it is quite obvious that, if t goes
Volta X3d X2d X1d on increasing, more and more number of terms will have to
be calculated to find the solution. To avoid this problem, some
may prefer to use the short memory principle based on the
Let the pseudo-random 8-bit unsigned integers for R- fact that distant past of the variable y have little effect on the
Channel, G-Channel and B-Channel are denoted by XR , XG G-L derivative. But nevertheless, truncating the summation in
and XB respectively. Then in the above case, XR , XG and XB (14) will introduce error in the solution.
are generated as follows Again, any interpreted language like Python or Matlab is
XR = X1a ⊕ X3b ⊕ X2c ⊕ X3d , inefficient in running loops. So, instead of implementing the
XG = X2a ⊕ X1b ⊕ X1c ⊕ X2d , solution of (2), (5), (7) and (9) directly in Python, the costly
loops are implemented in C using Cython. This proved to be
XB = X3a ⊕ X2b ⊕ X3c ⊕ X1d . a great deal in increasing the speed of the solver.

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 12,2022 at 08:34:21 UTC from IEEE Xplore. Restrictions apply.
 2014 International Conference on Control, Instrumentation, Energy & Communication(CIEC) 586

(a) (b)
Fig. 7. Histogram of the original image.
Fig. 5. (a) Original image, and (b) Encrypted image

(a) (b) Fig. 8. Histogram of the encrypted image.

Fig. 6. (a) Correctly decrypted image, and (b) Incorrectly decrypted image.

intensity-levels. This also shows the algorithm’s invulnerability

The encryption–decryption process is applied on the stan- against statistical attacks.
dard 512 × 512 colour Lena image. Fig. 5a shows the original Now, to check the autocorrelation between the pixel in-
image; Fig. 5b shows the encrypted image; Fig. 6a shows the tensities of adjacent pixels, autocorrelation diagrams for the
correctly decrypted image when all the initial conditions and original image as well as the encrypted image for few ran-
the derivative orders are kept exactly equal to those in the domly chosen points are plotted. The original colour image
encryption process. Fig. 6b shows an example of a ‘nonsense’ possesses a very high degree of autocorrelation as seen from
decryption when just the initial value x1a (0) in Lorenz’s Fig. 9. On the other hand, it can be easily noticed from Fig.
system is changed to 0.10000000000001 from 0.1 keeping 10, that there exists almost nil autocorrelation between any
everything else the same. two adjoining pixels (vertically, horizontally and diagonally)
of the encrypted image. This indicates that in the encrypted
V. S TATISTICAL A NALYSIS AND V ULNERABILITY T ESTS image, an attempt of anticipating the intensity-level of any
The quality of the encryption algorithm is assessed statis- pixel depending on the knowledge of the intensity-levels of
tically in the following subsections. other pixels, will simply go in vain. No useful information

A. Analysis of Key Sensitivity and Key space

As already mentioned, the key space for this encryption
algorithm is as huge as 10336 . A minute change in the control
parameters, thus results into a decrypted outcome which is
absolutely nonsense as in Fig. 6b. So, the key space is easily
verified to have very high sensitivity to changes in the control
parameters.

B. Image Histogram and Autocorrelation Analysis

The histograms of the original colour image and the
encrypted image are depicted in Fig. 7 and Fig. 8 respectively.
After observing these histograms, it can be inferred that the
encrypted image, generated by the proposed encryption algo- Fig. 9. Autocorrelation diagram of the original image for few randomly
rithm, contains almost equal number of pixels for all possible chosen points.

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 12,2022 at 08:34:21 UTC from IEEE Xplore. Restrictions apply.
 2014 International Conference on Control, Instrumentation, Energy & Communication(CIEC) 587

250
VI. C ONCLUSIONS
200 The use of multiple fractional order systems for image
Intensity of adjacent pixel encryption significantly increases the key-space and the key
150
sensitivity. The speed, and invulnerability of the algorithm
against brute-force attacks, make it useful for many practical
100
applications. Although the programme is optimised to run as
fast as possible, it will still seem sluggish for images of very
large size. Moreover, it should be kept in mind that, although
50
the derivative orders in the differential equations are considered
to be a part of the control parameters (the password), improper
0 choice of values for them may not satisfy the necessary
0 50 100 150 200 250
Pixel intensity
condition of the existence of double scroll attractors [11]. This
Fig. 10. Autocorrelation diagram of the encrypted image few randomly might fail the algorithm. So, proper restrictions on them must
chosen points. be made before executing the programmes.

R EFERENCES
can be gathered without knowing the exact keys.
[1] R. A. J. Matthews, “On the derivation of a chaotic encryption algorithm,”
Neural Computation, vol. 12, pp. 2519-2535, 2000.
C. Image Entropy [2] F. Y. Sun, S. T. Liu, Z. Q. Li, et al, A novel image encryption scheme
based on spatial chaos map, Chaos, Solitons, Fractals, Vol. 38, pp. 631-
Entropy of a single-channel image is expressed as [10] 640, 2008.
255
 [3] X. J. Tong, M. G. Cui, “Image encryption with compound chaotic
H=− p(l) log2 p(l), (15) sequence cipher shifting dynamically”, Image and Vision Computing ,
Vol. 26, pp. 843-850, 2008.
l=0
[4] H. H. Nien, W. T. Huang, C. M. Hung, S. C. Chen, S. Y. Wu, C. K. Huang
where, l stands for the intensity-levels varying between 0 and Y. H. Hsu, “Hybrid Image Encryption Using Multi-Chaos-System,”
and 255 and p(l) denotes the number of pixels for a certain in IEEE Conf. Information, Communications and Signal Processing, pp.
1-5, 2009.
intensity level l. Now, the entropy value of an image is zero
when it only incorporates pixels of an invariable intensity-level, [5] Ming Yin and Liwei Wang, “A new study in encryption base on fractional
order chaotic system ,” in Int. Conf. on Communications, Circuits and
and H has a maximum value of 8 when for any pixel location, Systems, 2008, pp. 637-640.
all the possible intensity levels are equally probable. Hence it is [6] A. G. Radwan, S. K. Abd-El-Hafiz and S. H. AbdElHaleem, “Image
crystal clear that a low entropy value refers to an inferior level encryption in the fractional-order domain ,” in Int. Conf. on Engineering
of ‘confusion’. For the original and the encrypted image, the and Technology, 2012, pp. 1-6.
following entropy levels for different channels are found. Thus, [7] I. Petráš, Fractional-Order Nonlinear Systems: Modeling, Analysis and
Simulation, Berlin Heidelberg: Springer-Verlag, 2011, pp. 103-184.
TABLE II. I MAGE E NTROPY [8] D. Valério and J. Sà da Costa, “Introduction to single-input, single-output
fractional control,” IET Control Theory Appl., Vol. 5, pp. 1033-1057,
Image under Channels
2011.
consideration R-Channel G-Channel B-Channel
Original Lena image 7.2596 7.5890 6.9734 [9] C. A. Monje, Y-Q Chen, B. M. Vinagre, D. Xue and V. Feliu, Fractional-
Encrypted image 7.6985 7.7213 7.6745 order Systems and Controls: Fundamentals and Applications, London:
Springer-Verlag, 2010, pp. 12-16.
[10] R. C. Gonzalez and R. E. Woods, Digital Image Processing, 3rd ed.
it can easily be seen that the encryption process significantly New Jersey: Prentice Hall, 2007, pp. 531-533.
increases the entropy levels for different channels. So, the [11] M. S. Tavazoei, and M. Haeri, “A necessary condition for double scroll
encrypted image is ‘more confusing’ than the original one. attractor existence in fractional - order systems,” Physics Letters A, Vol.
367, pp. 102-113, 2007.
This is also an indication that the algorithm is of sufficiently
good quality.

Authorized licensed use limited to: Indian Statistical Institute Kolkata. Downloaded on December 12,2022 at 08:34:21 UTC from IEEE Xplore. Restrictions apply.