Optics & Laser Technology 62 (2014) 152–160

Contents lists available at ScienceDirect

Optics & Laser Technology

journal homepage: www.elsevier.com/locate/optlastec

Novel image compression–encryption hybrid algorithm based

on key-controlled measurement matrix in compressive sensing
Nanrun Zhou a,b,n, Aidi Zhang a, Fen Zheng a, Lihua Gong a,c
a
Department of Electronic Information Engineering, Nanchang University, Nanchang 330031, China
b
Information Security Center, Beijing University of Posts and Telecommunications, Beijing 100876, China
c
Jiangxi Province Key Laboratory of Image Processing and Pattern Recognition, Nanchang Hangkong University, Nanchang 330063, China

art ic l e i nf o a b s t r a c t

Article history: The existing ways to encrypt images based on compressive sensing usually treat the whole measurement
Received 16 January 2014 matrix as the key, which renders the key too large to distribute and memorize or store. To solve this
Received in revised form problem, a new image compression–encryption hybrid algorithm is proposed to realize compression and
21 February 2014
encryption simultaneously, where the key is easily distributed, stored or memorized. The input image is
Accepted 25 February 2014
divided into 4 blocks to compress and encrypt, then the pixels of the two adjacent blocks are exchanged
Available online 31 March 2014
randomly by random matrices. The measurement matrices in compressive sensing are constructed by
Keywords: utilizing the circulant matrices and controlling the original row vectors of the circulant matrices with
Image encryption logistic map. And the random matrices used in random pixel exchanging are bound with the
Image compression
measurement matrices. Simulation results verify the effectiveness, security of the proposed algorithm
Compressive sensing
and the acceptable compression performance.
& 2014 Elsevier Ltd. All rights reserved.

1. Introduction
With the development of multimedia technology, more and
more information comes from images. The security of images
becomes a serious issue and hence a number of image encryption
algorithms were proposed [1–10]. For example, Chen et al. pro-
posed a new phase retrieval algorithm for optical image encryp-
tion in three-dimensional space [1], where the two-dimensional
plaintext was considered as a series of particles distributed in 3D
space, and an iterative phase retrieval algorithm was developed to
encrypt the series of particles into phase-only masks. Later, they
proposed a new optical image encryption method based on
multiple-region plaintext and phase retrieval in 3D space [2],
where the plaintext was divided into multiple regions and each
region was encrypted into one phase-only mask based on phase
retrieval in 3D space. An optical image encryption based on
coherent diffractive imaging using multiple wavelengths was
proposed [3], where the coherent diffractive imaging with multiple
wavelengths was applied into optical image encryption. He et al.
analyzed the collision property of the optical image encryption
technique based on interference [4] and found that various distinct
pairs of phase-only masks yielded almost the same outputs by use of
a modified phase retrieval algorithm. A multiple-image encryption
n
Corresponding author at: Department of Electronic Information Engineering,
Nanchang University, No. 999, Xuefu Avenue, Honggutan Xinqu, Nanchang 330031,
China. Tel.: þ86 791 83969670.
E-mail address: nrzhou@ncu.edu.cn (N. Zhou).
scheme was proposed based on the phase retrieval process and
phase mask multiplexing in the fractional Fourier transform domain
by Sui [5], where each original image was encoded into a phase-only
function by using the proposed phase retrieval process and all the
obtained phase functions were modulated into an interim, which
was encrypted into the final ciphertext by using the fractional
Fourier transform. Lu presented a novel method for optical image
encryption based on a modified radial shearing interferometer [6],
where the plaintext image was first encoded into a phase-only mask
and then modulated by a random phase mask; the result was
regarded as the input of the radial shearing interferometer and was
divided into two coherent lights, which interfered with each other,
leading to an interferogram. An optical authentication technique
based on interference image hiding system and phase-only correla-
tion was proposed by Yuan [7], where some predefined complex
images with different amplitudes and the same phase were respec-
tively encoded into two phase-only masks according to the inter-
ference principle. Their proposed technique can easily generate
different verification keys for different users, so it brought conve-
nience for multi-user application. A flexible multiple-image encryp-
tion algorithm based on log-polar transform and double random
phase encoding technique was proposed [8], where images were
transformed by log-polar transform and compounded, and then
encrypted by double random phase encoding. A color image
encryption algorithm was designed with the affine transform in
the gyrator transform domains [9], where the RGB components of
the color image were converted into the real part and the imaginary
part of a complex function by employing the affine transform and

http://dx.doi.org/10.1016/j.optlastec.2014.02.015
0030-3992/& 2014 Elsevier Ltd. All rights reserved.
 N. Zhou et al. / Optics & Laser Technology 62 (2014) 152–160
subsequently the complex function was encoded and transformed
in the gyrator domain. And Liu et al. proposed a double image
encryption scheme by using random phase encoding and pixel
exchanging in the gyrator transform domains, where the random
pixel exchanging (RPE) was introduced [10].
Compressive sensing (CS) [11,12] is a newly sampling–recon-
struction technique which can complete the sampling and com-
pressing simultaneously. Some researchers investigated the
security of image encryption algorithms with CS. Rachlin and
Baron investigated the security when eavesdroppers had no idea
of the measurement matrix and demonstrated a computational
notion of secrecy [13]. Orsdemir and Altun examined the security
and robustness of a compressive sensing based encryption algo-
rithm and indicated that the CS based encryption is computation-
ally secure [14]. Abdulghani and Rodriguze-Villegas showed the
additional benefits of compressive sensing in preserving data
privacy and indicated that the inherent multidimensional projec-
tion perturbation feature made it hard to breach the privacy [15].
And several image encryption algorithms based on CS have been
proposed. For example, compressive sensing was introduced in an
image encryption method based on double random-phase encod-
ing [16] to lower the encryption data volume due to the dimen-
sional decrease properties of CS [17]. Based on the method in [17],
the Arnold transforming was introduced later to enhance the
security [18]. Huang and Sakurai divided the original image to
blocks and vectorized each block to one-dimensional vectors, and
then encrypted and compressed these vectors with CS and block
Arnold scrambling [19]. Zhang and Ren proposed a scheme of
compressing and decompressing encrypted image based on CS
where the original image was encrypted by a secret orthogonal
transform and then compressed by CS with a pseudo-random
measurement matrix, and stated that the smoother the original
image, the better the quality of the reconstructed image [20]. To
overcome the problem that the measurement data from linear
dimension reduction projection directly serving as the encrypted
image failed to resist against the chosen-plaintext attack [21],
Huang et al. proposed a parallel image encryption method based
on CS, where block cipher structure consisting of scrambling,
mixing, S-box and chaotic lattice XOR is designed to further
encrypt the quantized measurement data [22]. Sreedhanya and
Soman employed both compressive sensing and Arnold scram-
bling to encrypt color image [23]. Athira et al. proposed a novel
image encryption algorithm based on CS, which generates the key
by linear feedback shift register [24].
While all the above CS-based encryption algorithms adopted
the whole measurement matrix as key, which renders the key too
large to distribute and memorize. The compression and the
encryption in some schemes cannot perform simultaneously. To
overcome these shortcomings, we explore a new hybrid compres-
sion–encryption algorithm where the measurement matrix is
controlled by keys and constructed as a circulant matrix. The plain
image is divided into 4 blocks to compress and encrypt, and then
the 4 compressed and encrypted blocks are scrambled by random
pixel exchanging with the random matrices.
The outline of this paper is as follows: some fundamental
knowledge is introduced in Section 2, the proposed algorithm is
described in Section 3, experimental results and discussion are
given in Section 4, and a brief conclusion is arrived at in Section 5.
2. Fundamental knowledge
2.1. Compressive sensing
The CS theory takes the space structure of signal in considera-
tion and samples signal in the space domain. In this way, CS does
153
not sample so many redundancy data. CS theory shows that a
signal x with length N has the representation in the Ψ domain:
α ¼ ΨTx ð1Þ
Projecting α onto a measurement matrix Φ of size M  N, one
can obtain an M  1 vector y, where M is the number of
measurements, and M⪡N, i.e.,
y ¼ Φx ¼ ΦΨ α ¼ Θα; ð2Þ
where the sensor matrix Θ is the product of Φ and Ψ , which
satisfies restricted isometry property (RIP) [11].
Definition of RIP: for each integer k ¼ 1; 2; …, define the
isometry constant δk of a matrix Φ as the smallest number such
that
ð1  δk Þ J f J 22 r J Φf J 22 r ð1 þ δk Þ J f J 22 ð3Þ
n
holds for all vectors f A R .
It is clear that the substance of RIP is that matrix Θ satisfied RIP
can keep the approximate Euclidean distance of k-sparse signal,
which ensures the k-sparse signal is not in the null space of Θ so
that it is possible to reconstruct the signal.
It is required to estimate the sparest solution to y ¼ Θα to
recover the signal x. The problem of estimating the sparse solution
can be expressed as
min J α J 0 subject to y ¼ Θα: ð4Þ
The above problem may be solved by exhaustive combinatorial
search. But it will become an NP-hard problem [25] for large N. To
overcome this problem, some reconstruction algorithms such as
matching pursuit (MP) [26], orthogonal matching pursuit (OMP)
[27] and smooth l0 algorithm (SL0) [28] and so on, have been
developed. SL0 is adopted in our proposed algorithm.
2.2. Random pixel exchanging
The random pixel exchanging process is shown in Fig. 1 [10].
I 1 and I 2 are two matrices. The variables m and n are the indices of
the matrices. R is a random matrix whose elements are all limited
in the interval ½0; 1. The symbol is defined to swap the pixels
0 0
at the left side and the right one. I 1 and I 2 are the outputs of the
pixel exchanging operation. The new position ðm0 ; n0 Þ is computed
as [10]
(
m0 ¼ f 1 ðm; nÞ ¼ 1 þ round fðM  1Þ sin ½πRðm; nÞg
; 1 r mr M; 1 r n r N
n0 ¼ f 2 ðm; nÞ ¼ 1 þ round ½ðN  1ÞRðm; nÞ
ð5Þ
where M and N are the sizes of the random matrix R. The matrices
I 1 and I 2 have M  N pixels. The round function is toward nearest
integer for input number. The mean value R of random matrix R
Fig. 1. The random pixel exchanging.
154 N. Zhou et al. / Optics & Laser Technology 62 (2014) 152–160
is [10]
1
R¼ ∑ Rðm; nÞ
M  N 8 m;n
If Rðm; nÞ 4 R, the pixels at the positions ðm; nÞ and ðm0 ; n0 Þ are
exchanged with each other for the two matrices I 1 and I 2 . If
Rðm; nÞ o R, the pixel exchanging is made in the inner pixels of
every matrix.
The inverse process of the RPE operation is implemented
according to Fig. 1 with an opposite sequence.
2.3. Logistic map
Since its pseudo-randomness and sensibility to the initial
condition, chaos system is often used in cryptography. The defini-
tion of logistic map is
xn þ 1 ¼ μxn ð1  xn Þ; xn A ð0; 1Þ:
It becomes chaotic when μ A ½3:57; 4.
3. Image compression–encryption hybrid algorithm based on
key-controlled measurement matrix
The measurement matrix Φ is constructed as a circulant matrix.
The original row vector of the circulant matrix is controlled by the
logistic chaos map. Suppose the size of Φ is M  N. The steps are as
follows:
Step 1: Generate a sequence with length 2N by logistic map
with initial condition x01 ; abandon the preceding N elements to
obtain the sequence s ¼ ½s1 ; s2 ; …; sN , which are used as the
original row vectors of the circulant matrices.
Step 2: The circulant matrix Φ is constructed with the original
row vectors s ¼ ½s1 ; s2 ; …; sN . To reduce the relevance among
the column vectors, the first element of vector Φði; 1Þ will be the
result of multiplying Φði 1; NÞ by λ, where 2 r i rM and λ 4 1,
and the iteration:
Φði; 1Þ ¼ λ U Φði  1; NÞ
Φði; 2 : NÞ ¼ Φði  1; 1 : N 1Þ
The proposed algorithm is fit for the image whose width equals
to height. Suppose the size of input image is N  N. The proposed
method is shown in Fig. 2, and the image encryption–compression
steps are as follows:
Step 1: Divide the input image into 4 blocks, block1, block2,
block3 and block4, as shown in Fig. 3, thus each block is of the
size ðN=2  N=2Þ.
Fig. 2. The process of the proposed encryption algorithm.
ð6Þ
Fig. 3. The method of dividing the image and the order of RPE.
Step 2: Construct two M  N=2 measurement matrices, Φ1 and
Φ2 , with keys x01 and x02 , and measure the block1 and the
ð7Þ
block3 with Φ1 , measure the block2 and block4 with Φ2 . C 1 , C 2 ,
C 3 and C 4 are the measurements corresponding to block1,
block2, block3 and block4.
Step 3: Scramble the adjacent blocks with the matrices R1 and
R2 in the order, ➀, ➁, ➂, and ➃. Since both the ranges of R1 and
R2 are ½0; 1, the ranges of measurement matrices Φ1 and Φ2 are
½0; λ, and the sizes of measurement matrices are the same as
R1 and R2 , set R1 ¼ Φ1 =λ, R2 ¼ Φ2 =λ. As shown in Fig. 3, the
pixels exchanging between block1 and block2, block3 and
block4 are guided by R1 , and the pixels exchanging between
block2 and block3, block4 and block1 are guided by R2 . C is the
encrypted image.
In the decryption and decompression process, the inverse
scrambling operation should be performed firstly. Notice that the
inverse scrambling order is the inverse order of scrambling order,
i.e., ➃, ➂, ➁, and ➀. Then one can obtain the decrypted and
decompressed blocks with the SL0 algorithm. The decrypted image
is the combination of these blocks.
4. Experimental results and analyses
ð8Þ
The gray image ‘Lena’ with resolution 256  256, which is
shown in Fig. 4, is served as the plain image in experiment. Thus
the size of each block is 128  128. Without loss of generality, we
adopt the DCT as basis matrix Ψ.
The parameters in experiments are x01 ¼ 0:11, x02 ¼ 0:23,
μ ¼ 3:99, λ ¼ 2, and M ¼ ð3=4  N=2Þ ¼ 96. The encrypted images
and the decrypted images are shown in Fig. 4. Here the compres-
sion ratio is 4/3, and the key length is 2. While the key length
would be M  N if the whole measurement matrices are the keys.
For example, for the same test image, the method in [17] treats the
whole measurement matrix as the key and the size of measure-
ment matrix is 192  256. That is to say, our proposed method
shortens the key greatly compared with the methods treating the
whole measurement matrix as the key.
On the other hand, the computational complexities of the other
parts are considered, since the computational complexities of CS in
all image compression–encryption algorithms involving CS are
similar to each other. There are 704 times float number multiply
operations, 512 times float number add operations, 192 times float
number shift operations in generating two measurement matrices
process, 256  256 times swap operations and 256  256 times
comparison operations in the RPE process. While in the experi-
ment of [17], the plain image is encrypted and compressed to the
size 192  256 firstly, then two random phase masks are necessary.
Assume two random phase masks are generated by logistic map,
 N. Zhou et al. / Optics & Laser Technology 62 (2014) 152–160 155

Fig. 4. (a) Lena; (b) encrypted image; and (c) correct decrypted image.

Fig. 5. Histogram: (a1) Lena; (a2) encrypted Lena; (b1) Cameraman; (b2) encrypted Cameraman; (c1) Peppers; and (c2) encrypted Peppers.
156 N. Zhou et al. / Optics & Laser Technology 62 (2014) 152–160

Fig. 6. Correlation distribution of two horizontally adjacent pixels in (a1) original Lena; (a2) encrypted Lena; (b1) original Cameraman; (b2) encrypted Cameraman;
(c1) original Peppers; and (c2) encrypted Peppers.

Table 1 there are 768  256 times float number multiply operations and
Correlation coefficients of adjacent pixels. 384  256 times float number add operations to generate two
random phase masks. One can see that the amount of computation
Correlation coefficient Horizontal Vertical Diagonal
of the proposed algorithm is much less.
Original Lena 0.9590 0.9217 0.9071
Encrypted Lena 0.0846 0.0583 0.0931 4.1. Histogram
Original Cameraman 0.9585 0.9346 0.9063
Encrypted Cameraman 0.0639 0.0539 0.0848 The image histogram is often used to analyze the performance of
Original Peppers 0.9529 0.9465 0.9150
Encrypted Peppers 0.0787 0.0582 0.0873
the image encryption algorithm. It is the best when the values in the
histogram of the encrypted image are fairly uniform in distribution,
 N. Zhou et al. / Optics & Laser Technology 62 (2014) 152–160 157

or the second best when the histograms of different original images horizontal, vertical and diagonal directions, 16,000 adjacent pixel
are similar to each other. Fig. 5(a1), (b1) and (c1) are the histograms pairs are selected randomly from original image and encrypted image,
of Lena, Cameraman and Peppers, respectively. And Fig. 5(a2), (b2) correspondingly. As shown in Fig. 6, the regular distribution reflects
and (c2) are the histograms of their encrypted images, correspond- the high correlation between two adjacent pixels in original image
ingly. The histograms of the three original images are obviously and the disordered distribution reflects the weak correlation between
different from each other, while their encrypted images have similar two adjacent pixels in the encrypted image. And the distributions of
histograms. After a large number of parallel experiments, we found vertical and diagonal directions share the similar modality.
that the histograms of the ciphertexts of different original images are The correlation coefficient is
similar to Fig. 5(a2), (b2) and (c2). That is to say, the proposed
algorithm can frustrate the statistical analysis attack. ∑N i ¼ 1 ðxi  xÞðyi  yÞ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
C ¼ s   ð9Þ
2 2
i ¼ 1 ðxi  xÞ
∑N ∑N
i ¼ 1 ðyi  yÞ ;
4.2. Correlation of two adjacent pixels

The correlation of two adjacent pixels in a meaningful image is where x ¼ 1=N ∑N i ¼ 1 xi and y ¼ 1=N ∑i ¼ 1 yi . The quantitative eva-
N

usually close to 1, while that of the encrypted image should be close to luation on correlation is compiled in Table 1. The correlation of the
0. To measure the correlations between two adjacent pixels in the plaintext is close to 1 in each direction, while the correlation of the

Fig. 7. Decrypted image with (a) Δ1 ¼ 10  16 ; (b) Δ2 ¼ 10  16 ; and MSE curves versus (c) x01 þ Δ1 ; and (d) x02 þ Δ2 .

Fig. 8. (a) Encrypted Lena using x01 ¼ 0:11 and x02 ¼ 0:23; (b) encrypted Lena using x01 þ Δ1 ¼ 0:11þ 10  16 and x02 ¼ 0:23; and (c) difference between two encrypted images
(a) and (b).
158 N. Zhou et al. / Optics & Laser Technology 62 (2014) 152–160

Fig. 9. The results of noise attacks with different noise strengths: (a) k ¼ 4; (b) k ¼ 8; (c) k ¼ 16; (d) k ¼ 32; and (e) MSE curve versus noise strength.

encrypted image is close to 0 in each direction. That is to say, demonstrate that the attackers cannot obtain useful information
the proposed algorithm removes the tight relationship between by statistical analysis and the proposed algorithm can resist statistical
adjacent pixels of the original image successfully. The results analysis.
 N. Zhou et al. / Optics & Laser Technology 62 (2014) 152–160 159

4.3. Key space Table 2

The PSNR for different compression ratios.
An image encryption algorithm should be secure even though Compressed
Original Compression Decrypted
everything is known except for the key. Thus a good encryption and encrypted PSNR(dB)
image ration image
image
algorithm should have a large enough key space. In the proposed
algorithm, x01 and x02 are used as keys. Here, the key space is
calculated for x01 as generate two different sequences χ and χ~ by 4:3 34.19
using x01 and x01 þ Δ as initial values and both sequences are of
length N, and define mean absolute error between the two
sequences as [29]
1 2:1 29.82
MAEðχ; χ~ Þ ¼ ∑ 9χ  χ~ 9 ð10Þ
N

The key space for x01 is equal to ð1=Δ0 Þ, where Δ0 is the value of
Δ for MAE ¼ 0. The simulation results show that Δ0 comes out to be 4:1 25.93
1  10  17 , i.e., the key space of x01 is 1  1017 . Similarly, the key
space of x02 is 1  1017 . Thus, the total key space is as large as 1034 .
If one wants to construct the correct measurement matrix by
exhausting the keys, she must calculate 1034 times which would 4:3 30.85
take much time. Thus the proposed algorithm is secure against
brute-force attack.

4.4. Key sensitivity analysis 2:1 26.71

A nice image encryption algorithm should be sensitive to the

keys, which means that a tiny change in the keys would cause a
great distortion in the decrypted image visually. Intuitively, the
4:1 22.64
proposed method is sensitive since the original row vectors of
circulant matrices are controlled by the logistic map. In order to
evaluate the key sensitivity, the mean square error (MSE) [30]
between decrypted image and original image is calculated as
1 identity standard deviation. If the encrypted image is contami-
MSE ¼ ∑½Iðx; yÞ  Dðx; yÞ2 ; ð11Þ nated by white Gaussian noise with different k, the corresponding
L  H x;y
decrypted images are shown as Fig. 9(a)–(d). The quality of
where L  H represents the total number of image pixels, Iðx; yÞ decrypted images does not degrade too much and the decrypted
and Dðx; yÞ denote the values of input image and output image at image from the contaminated encrypted image with k ¼ 16 is still
the pixel ðx; yÞ, respectively. Fig. 7(a) and (b) shows the wrong acceptable.
decrypted images with x01 þΔ1 and x02 þ Δ2 , respectively. The
decrypted images are distorted greatly and cannot show any 4.6. Compression performance
information about the original image visually. Fig. 7(c) and (d)
shows the MSE curves for x01 and x02 , respectively. The MSE is very The proposed algorithm can compress and encrypt the image
large with deviation 1  10  16 to x01 or 1  10  16 to x02 , and the simultaneously. To evaluate the quality of decrypted digital images
MSE is very small only when x01 and x02 are both correct. Thus the versus different compression ratios, peak-to-peak signal-to-noise
decrypted image can only be recognized if the keys are correct. ratio (PSNR) [17] is introduced as
The sensitivity can also be tested by comparing two encrypted
images obtained by using neighbor keys. Fig. 8(a) and (b) shows 2552
PSNR ¼ 10 log 2 2
ð13Þ
the encrypted images with x01 ¼ 0:11, x02 ¼ 0:23 and its neighbor ð1=N Þ∑N
i ¼ 1 ∑j ¼ 1 ½Rði; jÞ  Iði; jÞ
N

keys x01 ¼ 0:11 þ1  10  16 , x02 ¼ 0:23, respectively. And the dif-

ference between these two encrypted images is shown in Fig. 8(c).
The tiny change in the keys results in great changes in the
encrypted image.
From the above two cases, the proposed algorithm is sensitive
enough to the keys.
4.5. Noise attack
It is inevitable that the noise would directly impact the quality
of the decrypted image at the stage of image processing and image
transmission. Suppose the encrypted image is contaminated by
the noise as
C 0 ¼ C þ kG0
where C 0 and C are the noisy encrypted image and the unsullied
original encrypted image, respectively, k indicates the noise
strength, and G is the white Gaussian noise with zero-mean and
where Rði; jÞ and Iði; jÞ are the reconstructed image and the input
image, respectively. Table 2 lists the PSNR for different compres-
sion ratios. The quality of reconstructed image is good in some
degrees even if the compression ratio is 4 : 1, which means the
compression ability of the proposed method is nice, and this caters
to the requirement of compression prior to transmission. We
notice that the compression ability of the proposed method is
not as good as JPEG, while it would be improved with the
development of the reconstruction algorithms in CS.
5. Conclusion
We have demonstrated a new image compression–encryption
ð12Þ
hybrid algorithm based on compressive sensing and random pixel
exchanging, where the compression and the encryption are
completed simultaneously. The image is divided into 4 blocks to
compress and encrypt. Then random pixel exchanging is
160 N. Zhou et al. / Optics & Laser Technology 62 (2014) 152–160
introduced to scramble the compressed and encrypted blocks.
Compared with the methods adopting the whole measurement
matrix as key, the proposed algorithm shortens the key greatly,
which is of importance for a practical encryption algorithm. By
utilizing the circulant matrix to construct the measurement matrix
in CS and controlling the original row vector of circulant matrix
with chaos system, the proposed algorithm is secure. By introdu-
cing the random pixel exchanging and binding the random
matrices with the measurement matrices, the security is enhanced
further. The simulation results show that this image compression–
encryption hybrid algorithm can provide good security and nice
compression performance.
Acknowledgments
The authors would like to thank the editor and the anonymous
reviewers for their valuable suggestions and comments which
have led to a much improved paper. This work was supported by
the National Natural Science Foundation of China (Grant nos.
61262084 and 61141007); the Foundation for Young Scientists of
Jiangxi Province (Jinggang Star) (Grant no. 20122BCB23002), the
Natural Science Foundation of Jiangxi Province (Grant nos.
20114BAB201018 and 20132BAB201019); and the Opening Project
of Key Laboratory of Image Processing and Pattern Recognition
(Nanchang Hangkong University), Jiangxi Province (Grant no.
TX201204002).
References
[1] Chen W, Chen XD, Sheppard Colin JR. Optical image encryption based on
phase retrieval combined with three-dimensional particle-like distribution.
J Opt 2012;14:075402.
[2] Chen W, Chen XD. Optical image encryption based on multiple-region
plaintext and phase retrieval in three-dimensional. Opt Lasers Eng
2013;51:128–33.
[3] Chen W, Chen XD, Sheppard Colin JR. Optical image encryption based on
coherent diffractive imaging using multiple wavelengths. Opt Commun
2012;285:225–8.
[4] He WQ, Peng X, Meng XF. Collision in optical image encryption based on
interference and a method avoiding this security leak. Opt Lasers Technol
2013;47:31–6.
[5] Sui LS, Xin MT, Tian AL. Multiple-image encryption based on phase mask
multiplexing in fractional Fourier transform domain. Opt Lett 2013;38:1996–8.
[6] Lu DJ, He WQ, Peng X. Optical image encryption based on a radial shearing
interferometer. J Opt 2013;15:105405.
[7] Yuan S, Zhang T, Zhou X, Liu XM, Liu MT. Optical authentication technique
based on interference image hiding system and phase-only correlation. Opt
Commun 2013;304:129–35.
[8] Gong LH, Liu XB, Zheng F, Zhou NR. Flexible multiple-image encryption
algorithm based on log-polar transform and double random phase encoding
technique. J Mod Opt 2013;60:1074–82.
[9] Chen H, Du XP, Liu ZJ, Yang CW. Color image encryption based on affine
transform and gyrator transform. Opt Lasers Eng 2013;51:768–75.
[10] Liu ZJ, Zhang Y, Li S, Liu W, Liu WY, Wang YH, et al. Double image encryption
scheme by using random phase encoding and pixel exchanging in the gyrator
transform domains. Opt Laser Technol 2013;47:152–8.
[11] Donoho DL. Compressed sensing. IEEE Trans Inf Theory 2006;52:1289–360.
[12] Candès EJ. Compressive sampling. Int Congr Math 2006;3:1433–52.
[13] Rachlin Y, Baron D. The secrecy of compressed sensing measurements. In:
Proceedings of the 46th annual allerton conference on communication,
control, and computing; 2008. p. 813–17.
[14] Orsdemir A, Altun HO, Sharma G., et al. On the security and robustness of
encryption via compressed sensing. In: Proceedings of the IEEE military
communications conference MILCOM; 2008.
[15] Abdulghani AM, Rodriguze-Villegas E. Compressive sensing: from “compres-
sing while sampling” to “compressing and securing while sampling”. In:
Proceedings of the 32nd annual international conference of the IEEE engineer-
ing in medicine and biology society, EMBS'10; 2010. p. 1127–30.
[16] Philippe R, Bahram J. Optical image encryption based on input plane and
Fourier plane random encoding. Opt Lett 1995;20:767–9.
[17] Lu P, Xu ZY, Lu X, Liu XY. Digital image information encryption based on
compressive sensing and double random-phase encoding technique. Optik-Int
J Light Electron Opt 2013;124:2514–8.
[18] Liu XY, Cao YP, Lu P, Lu X, Li Y. Optical image encryption technique based on
compressed sensing and Arnold transformation. Optik-Int J Light Electron Opt
2013;124:6590–3.
[19] Huang R, Sakurai K. A robust and compression-combined digital image
encryption method based on compressive sensing. In: Proceedings of the
7th international conference on intelligent information hiding and multimedia
signal processing, IIHMSP; 2011. p. 105–8.
[20] Zhang XP, Ren YL, Feng GR, Qian ZX. Compressing encrypted image using
compressive sensing. In: Proceedings of the 7th international conference on
intelligent information hiding and multimedias signal processing, IIHMSP;
2011. p. 222–5.
[21] Yann F, Albertina C, Thomas JN, Bahram J. Resistance of the double random
phase encryption against various attacks. Opt Exp 2007;15:10253–65.
[22] Huang R, Rhee KH, Uchida S. A parallel image encryption method based on
compressive sensing. Multimed Tools Appl 2012:1–23.
[23] Sreedhanya AV, Soman KP. Secrecy of cryptography with compressed sensing.
In: Proceedings of the 2012 international conference on advances in comput-
ing and communications, ICACC; 2012. p. 207–10.
[24] Athira V, George SN, Deepthi PP. A novel encryption method based on
compressive sensing. In: Proceedings of the 2013 IEEE international multi
conference on automation, computing, control, communication and com-
pressed sensing, iMac4s; 2013. p. 271–5.
[25] Natarajan BK. Sparse approximate solutions to linear systems. SIAM J Comput
1995;24:227–34.
[26] Mallat SG, Zhang Z. Matching pursuits with time-frequency dictionaries. IEEE
Trans Signal Process 1993;41:3397–415.
[27] Liu E, Temlyakov VN. The orthogonal super greedy algorithm and application
in compressed sensing. IEEE Trans Inf Theory 2012;58:2040–7.
[28] Mohimani H, Babaie-Zadeh M, Jutten C. A fast approach for overcomplete
sparse decomposition based on smoothed l0 norm. IEEE Trans Signal Process
2009;57:289–301.
[29] Hennelly BM, Sheridan JT. Image encryption and the fractional Fourier trans-
form. Proc SPIE – Int Soc Opt Eng 2003;5202:76–87.
[30] Takanori N, Bahram J. Optical encryption system with a binary key code. Appl
Opt 2000;39:4783–7.