Citation 16 PDF

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/236188407
Assessing the performance of a method of simultaneous compression and

encryption of multiple images and its resistance against various attacks
Article in Optics Express · April 2013

DOI: 10.1364/OE.21.008025 · Source: PubMed
CITATIONS READS
53 161
4 authors, including:
Ayman Alfalou C. Brosseau

Institut Supérieur de l'Electronique et du Numérique Université de Bretagne Occidentale
169 PUBLICATIONS 1,867 CITATIONS 267 PUBLICATIONS 6,135 CITATIONS
SEE PROFILE SEE PROFILE
Maher Jridi
Institut Supérieur de l'Electronique et du Numérique
53 PUBLICATIONS 408 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
DSP power efficiency View project
Magnetic materials View project
All content following this page was uploaded by Ayman Alfalou on 20 May 2014.
The user has requested enhancement of the downloaded file.

Assessing the performance of a method of
simultaneous compression and encryption of
multiple images and its resistance against
various attacks
A. Alfalou,1,* C. Brosseau,2 N. Abdallah,1 and M. Jridi1
1
ISEN Brest, VISION- L@bISEN, 20 rue Cuirassé Bretagne, CS 42807, 29228 Brest Cedex 2, France
2
Université de Brest, Lab-STICC, CS 93837, 6 avenue Le Gorgeu, 29238 Brest Cedex 3, France
*
ayman.al-falou@isen.fr
Abstract: We introduce a double optimization procedure for spectrally

multiplexing multiple images. This technique is adapted from a recently
proposed optical setup implementing the discrete cosine transformation
(DCT). The new analysis technique is a combination of spectral fusion
based on the properties of DCT, specific spectral filtering, and quantization
of the remaining encoded frequencies using an optimal number of bits.
Spectrally multiplexing multiple images defines a first level of encryption.
A second level of encryption based on a real key image is used to reinforce
encryption. A set of numerical simulations and a comparison with the well
known JPEG (Joint Photographic Experts Group) image compression
standard have been carried out to demonstrate the improved performances
of this method. The focus here will differ from the method of simultaneous
fusion, compression, and encryption of multiple images (SFCE) [Opt.
Express 19, 24023 (2011)] in the following ways. Firstly, we shall be
concerned with optimizing the compression rate by adapting the size of the
spectral block to each target image and decreasing the number of bits
required to encode each block. This size adaptation is achieved by means of
the root-mean-square (RMS) time-frequency criterion. We found that this
size adaptation provides a good tradeoff between bandwidth of spectral
plane and number of reconstructed output images. Secondly, the encryption
rate is improved by using a real biometric key and randomly changing the
rotation angle of each block before spectral fusion. By using a real-valued
key image we have been able to increase the compression rate of 50% over
the original SFCE method. We provide numerical examples of the effects
for size, rotation, and shifting of DCT-blocks which play noteworthy roles
in the optimization of the bandwidth of the spectral plane. Inspection of the
results for different types of attack demonstrates the robustness of our
procedure.
©2013 Optical Society of America
OCIS codes: (070.0070) Fourier optics and signal processing; (100.0100) Image processing;
(200.4560) Optical data processing.
References and links

1. A. Alfalou and C. Brosseau, “Optical image compression and encryption methods,” Adv. Opt. Photon. 1(3), 589
(2009).
2. B. Javidi, ed., Optical and Digital Techniques for Information Security (Springer Verlag, New York, 2005).
3. P. Refregier and B. Javidi, “Optical image encryption based on input plane and Fourier plane random encoding,”
Opt. Lett. 20(7), 767–769 (1995).
4. Y. Frauel, A. Castro, T. J. Naughton, and B. Javidi, “Resistance of the double random phase encryption against
various attacks,” Opt. Express 15(16), 10253–10265 (2007).
#183032 - $15.00 USD Received 8 Jan 2013; revised 21 Feb 2013; accepted 28 Feb 2013; published 26 Mar 2013
(C) 2013 OSA 8 April 2013 / Vol. 21, No. 7 / OPTICS EXPRESS 8025
5. P. Kumar, J. Joseph, and K. Singh, “Known-plaintext attack-free double random phase-amplitude optical
encryption: vulnerability to impulse function attack,” J. Opt. 14(4), 045401 (2012).
6. J.-J. Huang, H.-E. Hwang, C.-Y. Chen, and C.-M. Chen, “Lensless multiple-image optical encryption based on
improved phase retrieval algorithm,” Appl. Opt. 51(13), 2388–2394 (2012).
7. X. Wang and D. Zhao, “Double images encryption method with resistance against the specific attack based on an
asymmetric algorithm,” Opt. Express 20(11), 11994–12003 (2012).
8. A. Alfalou and A. Mansour, “All optical video-image encryption enforced security level using independent
component analysis,” J. Opt. A, Pure Appl. Opt. 9, 787 (2007).
9. Z. Liu, Q. Guo, L. Xu, M. A. Ahmad, and S. Liu, “Double image encryption by using iterative random binary
encoding in gyrator domains,” Opt. Express 18(11), 12033–12043 (2010).
10. A. Alfalou and C. Brosseau, “Dual encryption scheme of images using polarized light,” Opt. Lett. 35(13), 2185–
2187 (2010).
11. F. Mosso, J. F. Barrera, M. Tebaldi, N. Bolognini, and R. Torroba, “All-optical encrypted movie,” Opt. Express
19(6), 5706–5712 (2011).
12. M. Paturzo, P. Memmolo, L. Miccio, A. Finizio, P. Ferraro, A. Tulino, and B. Javidi, “Numerical multiplexing
and demultiplexing of digital holographic information for remote reconstruction in amplitude and phase,” Opt.
Lett. 33(22), 2629–2631 (2008).
13. A. Alfalou and C. Brosseau, “Exploiting root-mean-square time-frequency structure for multiple-image optical
compression and encryption,” Opt. Lett. 35(11), 1914–1916 (2010).
14. E. Darakis and J. J. Soraghan, “Reconstruction domain compression of phase-shifting digital holograms,” Appl.
Opt. 46(3), 351–356 (2007).
15. T. J. Naughton, J. B. McDonald, and B. Javidi, “Efficient compression of fresnel fields for Internet transmission
of three-dimensional images,” Appl. Opt. 42(23), 4758–4764 (2003).
16. T. J. Naughton, Y. Frauel, B. Javidi, and E. Tajahuerce, “Compression of digital holograms for three-dimensional
object reconstruction and recognition,” Appl. Opt. 41(20), 4124–4132 (2002).
17. S. L. Wijaya, M. Savvides, and B. V. K. Vijaya Kumar, “Illumination-tolerant face verification of low-bit-rate
JPEG2000 wavelet images with advanced correlation filters for handheld devices,” Appl. Opt. 44(5), 655–665
(2005).
18. L. Ding, Y. Yan, Q. Xue, and G. Jin, “Wavelet packet compression for volume holographic image recognition,”
Opt. Commun. 216(1-3), 105–113 (2003).
19. S. Yeom, A. Stern, and B. Javidi, “Compression of 3D color integral images,” Opt. Express 12(8), 1632–1642
(2004).
20. A. Alfalou, C. Brosseau, N. Abdallah, and M. Jridi, “Simultaneous fusion, compression, and encryption of
multiple images,” Opt. Express 19(24), 24023–24029 (2011).
21. M. Johnson, P. Ishwar, V. Prabhakaran, D. Schonberg, and K. Ramchandran, “On compressing encrypted data,”
IEEE Trans. Signal Process. 52(10), 2992–3006 (2004).
22. A. Alfalou, A. Mansour, M. Elbouz, and C. Brosseau, “Optical compression scheme to multiplex &
simultaneously encode images”, in Optical and Digital Image Processing Fundamentals and Applications, G.
Cristobal (Ed.), P. Schelkens (Ed.), and H. Thienpont (Ed.), (Wiley, 2011) pp. 463–483.
23. J. Krommweh, “Tetrolet transform: A new adaptive Haar wavelet algorithm for sparse image representation,” J.
Vis. Commun. Image R. 21(4), 364–374 (2010).
1. Introduction
In the last decade optical encryption has emerged as a framework for studying information
processing [1–3]. However, it is well established that the standard double random phase
encryption (DRP) exhibits vulnerability to various attacks [4,5]. Several approaches have been
reported to improve the encryption efficiency of DRP, see, e.g. Refs [1–8]. One possible route
towards a multiple-image optical encryption method is based on the modified Gerchberg-
Saxton algorithm (MGSA) and the Fresnel transform [6]. Recently, Alfalou and Mansour [8]
found that an approach based on independent component analysis produces a significantly
enhanced resistance of the DRP method against specific attacks. In a different context,
polarized light and a gyrator transform are used to validate optical encryption methods [9,10].
More recently, Wang and co-workers [7] suggested an alternative double-image encryption
resistance method. Another encryption method for videos relying on DRP and a specific
merging procedure in the output plane was put forward by Mosso et al. [11]. It is worth
observing that this method leads to good performances even if its compression part is not
optimized. However, the method proposed by Mosso and co-workers does not consider the
issue of spectral overlapping in the output plane. Consequently it requires a large number of
bits to encode the output plane. In addition, their quantization procedure leads to a decrease of
performances when the images are reconstructed.
While a large body of work has been reported on encryption, the prevailing opinion is that
the compression stage is not a main concern in the field of optical encryption. Following
existing published work, the real and imaginary parts of the output image following DRP [1–
3] have been considered separately. On the other hand, many digital hologram (DH) studies
have been performed for decreasing the amount of data to be processed. For example, Paturzo
and co-workers [12] investigated the possibility to multiplexing and demultiplexing
numerically DHs with the aim of optimizing their storage and/or transmission process. The
holograms are multiplexed and demultiplexed thanks to the unique property of the DHs to
numerically manage the complex wavefields. They showed that it is possible to correctly
retrieve quantitative information about the amplitude and phase of one hundred DHs. Their
approach can be useful to transmit efficiently, in terms of reduced amount of data, the DHs
from the recording head to a remote display unit. Other recent papers dealing with
compression of DHs are worthy of note since the methods they describe can be implemented
optically and can be adapted to our issue of simultaneous compression and encryption. For
example, a compression and multiplexing architecture of multiple images was reported in
[13].This method, which is based upon a specific spectral multiplexing (fusion without
overlapping) of the multiple images, aims to achieve a single encrypted image, at the output
plane of our system, that contains all information needed to reconstruct the target images. For
that purpose, the Fourier plane of the image to transmit is divided into two types of area, i.e.,
specific and common areas to each target image. A segmentation criterion taking into account
the RMS criterion is proposed. This approach, which consists of merging the input target
images together (in the Fourier plane) allows us to reduce the information to be stored and/or
transmitted (compression) and induce noise on the output image (encryption). In the same
vein, Darakis and Soraghan [14] carried out a compression scheme of data, based on phase-
shifting interferometry digital holography (PSIDH) that is performed at the reconstruction
plane. These authors have shown how the increased spatial correlation apparent at the
reconstruction plane can be effectively exploited to obtain high compression even with
relatively simple methods such as the quantization followed by lossless coding. Naughton et
al. [15] compressed PSDH for the transmission of three dimensional images. For real-time
networking applications, the time required to compress can be as critical as the compression
rate. They achieved lossy compression through quantization of both the real and imaginary
streams, followed by a bit packing operation. They verified transmission speedup due to
compression using a special-purpose Internet-based networking application. In [16], Naughton
and co-workers also presented the results of applying lossless and lossy data compression to a
three-dimensional object reconstruction and recognition technique based on PSDH. The lossy
techniques are based on subsampling, quantization, and DCT. Dealing with a face recognition
application, Wijaya et al. [17] suggested using the image compression standard JPEG2000,
which is a wavelet-based compression engine used to compress the face images to low bit
rates suitable for transmission over low-bandwidth communication channels. Ding et al. [18]
also reported a wavelet-based method to recognize target images. In addition, Sekwom et al.
[19] proposed a compression method designed for color images based on standard MPEG-2.
As mentioned above, in contrast with optical encryption, optical compression techniques
have not received the amount of attention they deserve and were not mentioned in the above
mentioned analysis [1–11,20,21]. However, our ability to develop new technologies in
communication systems, secure local storage and network-transmission of images will largely
depend on our understanding of encryption and compression methods. Thus, combining
compression and encryption is important for most applications in communications. Further,
such approaches are often achieved either in cascade or independently. Motivated by these
observations and in anticipation of wide uses for optical communications, we set out to
investigate how one of these operations can affect the other.
1.1. Optical DCT: motivation and background
In Ref [13], we proposed a simultaneous compression and encryption method of multiple
images based on a spectral fusion of target images in the Fourier domain. Additionally,
Alfalou and Brosseau recently examined a new SFCE method by using DCT and information
fusion in the spectral domain, which is appropriate for dealing with multiple images of a video
sequence [20]. Two advantages of this method are noteworthy. Firstly, only real-valued
spectra are dealt with. Secondly, the use of DCT allows the grouping of the information for
reconstructing the image in the upper left corner of its spectral plane. Such a grouping of real-
valued information does not require the storage and/or transmission of complex-valued data as
is the case of most DRP-based techniques [1]. Previously, we have proposed and validated
[22] an all-optical setup which allows optical implementing of the DCT. However, the quality
of the reconstructed images remains poor because the data encoding uses a number of bits
which is not adapted to the dynamics in the spectral domain. The motivation of the present
study is to remedy this situation by introducing a detailed double optimization procedure for
spectrally multiplexing multiple images.
1.2. Optimization of the method
In this work we adopt the optimization strategy recently proposed in [20]. Its main advantage
in our context is that it offers a way to increase both compression (section 3) and encryption
(section 4) rates. Interestingly, the SFCE method has a strong ability to merge different
information from target images in the DCT spectral plane. Its underlying principle is to
decompose the spectral plane into many independent areas. One information block
corresponding to a target image is affected to each area. We follow the RMS time-frequency
criterion [13] introduced by Alfalou and Brosseau to determine the useful spectral size of
target images. As a consequence, the size of each block is adapted to the target image. We
wish to point out that this procedure differs from that which was considered by us in an earlier
work [20] where each block was set at a fixed size leading to degradation in performance for
image reconstruction. We found that this size adaptation provides a good tradeoff between
bandwidth of spectral plane and compression rate. In addition we include the idea of using a
real key for increasing the encryption rate of our method. This leads to a decrease of the
amount of encrypted information to store and/or to transmit compared to the SFCE method
which uses both real and imaginary components (section 4). The present results will allow us
to investigate the effects of quantization and several types of attack on compression rate.
The paper is organized as follows. First we briefly review the SFCE method by focusing
on the block selection and encryption methods. In section 3 we discuss the phenomenology of
the SFCE method in the context of increasing compression rates. A comparison of the error in
the reconstructed object following our procedure with others, e.g. JPEG image compression
standard, reveals its good performance. Section 4 is devoted to improvement of the encryption
performance. To test the robustness of our procedure and illustrate better the implications of
our findings, we simulate several types of attack. We summarize and conclude in section 5.
2. Simultaneous fusion, compression, and encryption of multiple images
Figure 1 (a) schematizes the basic principle of fusion underlying the SFCE method. A detailed
description of this method can be found in Ref [20]. Here, only the salient features relevant to
this analysis will be highlighted. In this figure, I(1), I(2), ..., I(Ni) are the target images at the
input of the SFCE system. Firstly we apply separately the DCT transform separately to each
of these images. Secondly, each spectrum of size (N,N) pixels is multiplied by a low-pass
filter ( N ' , N ' ) with N ' < N . In this way, a block of size ( N ' , N ' ) , containing the relevant
information for reconstructing each target image, is obtained. An example of such block
(filtered DCT) is shown in Fig. 1(b).
N
Filter (1)
DCT
I(1) X I(1)
First 4 filtered
Shift (1)
DCT DCT
I(2) X I(1)
Shift (2) ...
(a)
Filter (2)
...
...
.
.
.
2N'
N i
Filter (N )
Shift (N i )
DCT
I(Ni) X I(1)
{
N'
N' N'
Area (1)
Area (2)
(b)
.
.
.
Area (j)
filtered DCT filtered quantized DCT
Fig. 1. Principle of the SFCE method: (a) synoptic diagram, (b) compression technique scheme.
Each spectral plane is multiplied by a low-pass filter, of size set to ( N ' , N ' ) pixels, positioned
in its upper left corner. Then, four of these blocks to which are applied one of the four
rotations with angle 0°, 90°, 180°, 270° are grouped together. This leads to a domain of size
(2N’,2N’) pixels as shown in Fig. 1(a) (first 4 filtered DCTs). To avoid information overlap
these blocks are returned and shifted. Then, quantization to each block of size (N’,N’) pixels is
realized, as displayed in Fig. 1(b). This permits to divide each block in several areas which are
further quantized. This quantization is realized by normalizing each DCT to its maximum
value. Following this quantization, we get the block shown in the right-side of Fig. 1(b). By
reducing the maximum value in a given area we reduce the number of bits required to encode
it.
According to the above discussion, we have a real-valued spectral plane containing
relevant information for reconstructing the Ni target images. The use of a first encryption-key
(a specific image among those used in the input plane of SFCE), the specific fusion after
rotation and the quantification constitute the first level of encryption. Then, a digital
fingerprint encrypted with the classical DRP system is added to the fusion spectrum to avoid
the second encryption level.
Further, we note that even if the SFCE approach has proved to be an effective technique
[20], it has limitations because a simple low-pass filter of size (N’,N’) is insufficient to
optimize the bandwidth of the spectral plane. In addition, we remark that using a unique and
constant filter size for each block and image would not be optimized. It should also be noted
that the number of bits for encoding frequencies is not optimized when only the maximum
value of the given area is considered. In addition the encryption technique developed in [20]
requires the storage and/or transmission of complex data. As already stated in the Introduction
section, the problem we address now is to propose specific optimizations for each limitation
mentioned.
3. Compression optimization
Our optimization scheme begins with choosing the size of the filter used in the spectral plane
of each target image. In our proposed optimization scheme, we used and adapted the RMS
criterion to determine the minimal size for each target image.
3.1. Adaptation of the block size according RMS criterion
Optimizing the spectral bandwidth, that is increasing the multiplexing capability of the SFCE
method, requires finding a criterion which allows us to adapt the block size (set for the
moment to (N’,N’) pixels) for each target image. Let us estimate the minimum size of the
spectrum of a given target image. For that purpose we have adapted the RMS duration
criterion to DCT. Following [13], we define the RMS criterion
+∞ +∞
2
t=n   (u + v 2 ) S I (u, v ) dudv ,
2
(1)
−∞ −∞
where t is the block size expressed in pixels for a given target image, S I (u, v ) is the
normalized DCT of target image I, and n is a constant ranging from 0 and 2, e.g. n = 1
represents the minimum size required for reconstructing a given image.
Table 1. Example of a reconstructed image corresponding to several sizes of block (t , t )
(a)
Target image ( 256 × 256 ) pixels
(b) (c) (d)

Reconstructed image, n = 0.25 , Reconstructed image, n = 0.5 , Reconstructed image, n = 0.75 ,
( 22 × 22 ) pixels ( 44 × 44 ) pixels ( 66 × 66 ) pixels
(e) (g)
(f)
Reconstructed image, n = 1 , Reconstructed image, n = 2 ,
Reconstructed image, n = 1.5 ,
( 88 × 88 ) pixels (132 × 132 ) pixels (175 × 175 ) pixels
The value of 2 was achieved by trial and error and by considering that the useful
information contained in a DCT spectrum is contained in a space less than two times the
useful size defined by the RMS criterion. Hence, the filter size which is used in the spectral
plane (i.e. block’s size) is not constant but depends on the amount of information. We present
in Table 1 the reconstructed images corresponding to several sizes of block (t , t ) with n
ranging between 0 and 2. Table 1 shows that the reconstruction of the target image does not
require the information contained in all pixels. If n = 1, i.e. for a block of size 88 × 88 pixels
(panel (e) of Table 1), the algorithm was able to reconstruct satisfactorily the target image
(panel (a) of Table 1). However, the algorithm performs poorly for n<1 (panels (b-c-d) of
Table 1). In contrast a good performance can be achieved for n>1 (panels (f-g) of Table 1).
Changing n from 1.5 to 2 improves the quality of the reconstructed image. Beyond n = 2
there is no visible improvement of the quality of the reconstructed image. In the following n is
set to 2. For the purpose of later comparison of the numerically derived effect of block size on
the quality of the reconstructed image and compression rate, two different metrics will be
used: the mean-square error (MSE), and the peak signal-to-noise ratio (PSNR).
3.2. Effect of the block size on the quality of reconstructed images
Here no quantization is applied and the compression rate Tc _ pixel (expressed in pixels) is
evaluated as follows. Assume that we consider Ni target images of size (t x , t y ) pixels, where tx
and ty denote the size in pixels of each image (in this work we assume that t x = t y = N ). Then,
the total size of the image is N i t x t y = N i N 2 pixels. After multiplexing (Fig 1(a)), a single
spectral plane of size (t x , t y ) needs to be stored instead of the Ni images each of size (t x , t y )
pixels. Since the DCT spectrum is real, this size is equal to t x t y = N 2 As a result, we find that
size of multiplexed DCT spectral plane N2 1

Tc _ pixel = 1 − = 1− = 1− (2)
size of N i input images Ni × N 2
Ni
To illustrate this point, we consider Table 2 which shows a comparison of the reconstructed
images and compression rates as a function of block size: The first column of this table is
concerned with the block size obtained from Eq. (1). The second column corresponds to the
compression rates (in pixels) for a given block size, e.g. if N=256 (image size equal to 256 ×
256 pixels) and n=1.5, 4 blocks of size (132 × 132) pixels can be merged in a (256 × 256
pixels) spectral plane. The third column lists the MSE and PSNR values. The fourth column
shows an example of a reconstructed image. The fifth column shows the number of target
images that can be merged for a given n. Not surprisingly, the effect of increasing the number
of multiplexed images is to increase Tc _ pixel but significantly impacts the quality of the
reconstructed images. Increasing Tc _ pixel (while keeping a good quality of the reconstructed
images) can be realized by adapting quantization such as using the smallest number of bits.
Table 2. Effect of the compression rate ( Tc _ pixel ) in pixels on the quality of the
reconstructed images
Bloc size Compression rate MSE Reconstructed Number of

t (pixels) PSNR image multiplexed
images
n = 1.5 75% 4
MSE = 9.9 × 10−4
(132 × 132 ) PSNR = 30
n =1 83.3% 8
MSE = 1.4 × 10−3
( 85 × 85 ) PSNR = 28.7
n = 0.65 88.9% −3 16
MSE = 1.9 × 10
( 57 × 57 ) PSNR = 27.2
3.3. Calculation of the compression rate without quantization

For a further comparison of the effect of quantization on Tc, we first consider the special case
in the absence of quantization. The compression rate Tc (in bits) is determined with respect to
the input images (images to transmit and/or to store) and the filtered multiplexed spectra. It is
given by
Tc = 1 − sizeout size in , (3)
where size in = 8 N i t x t y is defined as the length (in its bit description) of the sequence of
images to store and/or transmit (gray scale images encoded over 8 bits are considered here),
and size out denotes its counterpart at the output of Fig. 1(a). The number of bits used in each
block can be written as log 2 {max ( block ) − min ( block )}t 2 , where max(block) and min
(block) denote respectively the maximal and minimal value in the blocks, and t is calculated
from Eq. (1).
It is interesting to point out that without taking into account the decimal part of the DCT
(without quantization) the encoding is realized with less than 16 bits compared with 65 bits
when the decimal part is considered. Without quantization Tc in bits is not optimized. That is,
it is smaller than Tc in pixels (Table 3). This is due in part to the rapidly varying spectra which
necessitate a large number of encoding bits. To remedy this situation we propose to quantify
the DCT blocks to obtain a fixed number of bits.
Table 3. Comparison of the compression rate expressed in bits and in pixels
Bloc size Compression rate (pixels) Compression rate

t (pixels) Tc_pixel Tc
n = 1.5 75% 51.5%
(132 × 132 )
n =1 83.3% 67.7%
( 85 × 85)
n = 0.65 88.9% 78.5%
( 57 × 57 )
3.4. Fixed quantization
A natural question is this: what is the influence of the number of bits m which is required to
encode the coefficients of the DCT blocks? There are several ways in which this question
might be answered, but here, we first assume that m=log2(max(block)) is fixed. The results
showing the reconstructed images, Tc, and PSNR are summarized in Table 4.
Table 4. Compression rate, PSNR, and quality of the reconstructed images when using a
number of bits set to m=log2(max(block))
Reconstructed
Input image Reconstructed image
image
Ni =2 Ni =4
Tc = Tc =
2.5% 51.4%
PSNR = 30.0
PSNR = 33.0
Reconstructed
Reconstructed image Reconstructed image
image
Ni =6 Ni =9 Ni=16
Tc = Tc = Tc =
67.7% 78.5% 88.0%
PSNR = 28.7 PSNR = 27.1 PSNR = 23.3
A notable feature of Table 4 is that obtaining a good compression rate requires a small block t,
i.e. a large number of images to be multiplexed. According to these numbers, the quality of
the reconstructed images degrades rapidly. For example, if Ni is set to 16, the attained Tc=88%
corresponding to a 10 dB loss. In addition, one sees that the values of Tc remain quite limited.
Increasing Tc necessitates optimizing the quantization in order to decrease the DCT values.
To attain this goal, we first address the case of fixed quantization. In [20], Tc was calculated in
terms of the actual values in each area. Hence it cannot be adapted according the specificities
of an application. As mentioned above we will quantify all values of every block with a fixed
number of bits, i.e. m. Hence, the block (t,t) resulting from this quantization necessitates a
reduced number of bits. According this analysis, the new values of DCT with quantization are
defined as follows
 ( 2m −1 − 1)Vblock ( i, j ) 
'
Vblock ( i, j ) = round  , (4)
max ( block ) 
 
where max(block ) denotes the maximum value of DCT in the given block, Vblock (i, j ) is the
DCT value before quantization for a given pixel (i,j), and m is the number of bits for a given
value of Tc. Note that in our previous work [20] several maximum values were calculated in a
given block according the quantized area.
Table 5. Effect of uniform quantization of the DCT blocks on the compression rate and
reconstructed image quality
Reconstructed
Reconstructed image Reconstructed image
image
Ni = 2 m Ni = 2 m Ni = 2
= 15 =8 m=4
Tc = Tc = Tc =
11.0% 54.8% 81.2%
PSNR=3 PSNR= PSNR
3.0 21.7 =15.9
Ni = 4 m Ni = 4 m Ni = 4
= 15 =8 m=4
Tc = Tc =
Tc= 90.3%
55.5% 77.9%
PSNR=3 PSNR= PSNR
0.0 21.7 =15.8
Ni = 6 m Ni = 6 m Ni = 6,
= 15 =8 m=4
Tc = Tc = Tc =
70.7% 85.7% 93.5%
PSNR=2 PSNR= PSNR
8.70 21.6 =15.5
Ni = 9 m Ni = 9 m Ni = 9
= 15 =8 m=4
Tc = Tc = Tc =
80.2% 90.6% 95.3%
PSNR=2 PSNR= PSNR
7.2 21.5 =15.4
Ni =16 Ni =16 Ni = 16 m
m =15 m=8 =4
Tc = Tc = Tc =
88.3% 94.0% 97.6%
PSNR=2 PSNR= PSNR
2.8 20.4 =15.2
The data in Table 5 show that compression rate can be fixed by choosing the number of
bits which is necessary to encode the values of each block. This corresponds to a compression
rate of 80% for a number of bits set to 4 (with only two images Ni=2; second line). This value
of Tc is significantly larger than the 2.5% obtained for the case of two images (Table 4). But, it
should be noted that the quality of the reconstructed images remains poor. Thus, if the latter is
over 8 bits this method leads to a high degradation in reconstructed image quality due to the
uniform quantization.
Comparing Table 4 (good quality of reconstructed images) and Table 5 (large Tcs)
indicates that a compromise between the two ways of calculating m should be found. We
suggest that a convenient measure of m is given by
m = log 2 (V ' ) , (5)
where m is the maximum number of bits used for encoding the DCT coefficients of a given
block, and V ' is the normalized value of the DCT coefficient within the range [0, max]. It is
worth noting that Eq. (5) can be applied to every block size ( t × t ) , and even to areas smaller
than ( t × t ) as depicted in Fig. 1(b).
3.5. Quantization by block optimization and normalization between 0 and max

To overcome the issue of image quality (Table 5), we adopt the following strategy. We
decompose a DCT block into several areas, and quantize each area using to Eq. (4). However,
the number of bits is not optimized. This reflects the fact that only the maximum value of the
quantified-DCT is considered in Eq. (4) without knowledge of its minimum value. The
optimization which is proposed consists in calculating a new block of the quantized DCT
taking into account of the minimum value (Fig. 1). The number of bits which is used to
encode each area of the DCT blocks is fixed and set to m. The DCT is first divided in several
areas according to a given step (Fig. 1). Next, each area is quantized (each area is represented
by a specific color in Fig. 1 by considering both maximum and minimum values in this area.
Table 6 compares the quality of the reconstructed images and Tc with or without
optimization. From left to right, the first column indicates the number of images to be
multiplexed in the spectral plane. PSNR is shown in the second column. The reconstructed
images without optimization [20] are displayed in the third column, while those with
optimization are considered in the fourth column. In this column, the number of bits used for
each area of the DCT blocks of size (t,t) is calculated with respect to the range between these
two values, i.e. log2(max-min), instead of using the range from 0 to the maximum value, i.e.
log2(max-0). The net effect is a decrease of m for the following reason. Assume that in a given
area the maximum and minimum value as are respectively set to 255 and 128. Using
log2(max-0) requires 8 bits for encoding these values. Now normalizing values between 0 and
max [0-128] and using log2(max-min) involves only 7 bits. Another notable feature of Table 6
is that as the PSNR is decreased the quality of the reconstructed image decreases gently while
the compression rate (in bits) increases. For a given PSNR value, the compression rate is
always larger when our optimization method is employed. This scheme permits a compression
rate close to 97% to be obtained with good quality of reconstructed images (Ni=16).
Table 6. Comparison of the compression rate and PSNR with or without optimization of
the number of bits m
Number of PSNR Reconstructed image Reconstructed image

multiplexed Tc [20] Tc
images (current work)
2 33
Tc = 9.0% Tc=15.0%
2 30.5
Tc = 72.9% Tc = 74.6%
2 18.4
Tc = 87.4% Tc = 89.0%
6 28.7
Tc = 70.8% Tc = 75.9%
6 27.6
Tc = 90.9% Tc = 92.0%
6 19.4
Tc = 93.6% Tc = 96.5%
16 23.3
Tc = 89.3% Tc = 90.8%
16 23.1
Tc = 95.3% Tc = 97.3%
16 18.3
Tc = 97.2% Tc = 99.1%
3.6. Comparison with JPEG
The efficiency of our optimized method will be demonstrated by adapting a number of pixels
to each image and DCT quantization in order to reduce the number of bits required for
encoding DCT blocks (t,t). Table 7 shows a comparison of the results (gray-level images)
obtained from our method with those obtained with the JPEG image compression standard for
several compression rates. The data in Table 7 shows that the optimized SFCE method proved
to be a very effective technique to control Tc. This is due to the adapted quantization and
choice of the size of the DCT block. The key feature is that our technique does not suffer from
the pixellization as is observed with the JPEG compression method. It is anticipated that this
can be improved further by applying different geometric forms to define each DCT block as
used in the Tetrolet transform [23]. Overall, these simulations confirm the importance of the
optimized SFCE method. As shown in Table 5, for a compression rate of ≈ 94% , the attained
PSNR is ≈ 27.7 , while for a compression rate of ≈ 93% JPEG leads to a PSNR value of
≈ 25.6% .
In Fig. 2(a) and Fig. 2(b) we present respectively the compression rate and the PSNR
values as a function of the number of multiplexed images without optimization. Figure 2(c)
compares the PSNR as a function of compression rate for the optimized SFCE and JPEG
methods. With the former, we get Tc=62% and PSNR= 32.5 dB.
100 %
32
80 % 29 dB
Compression rate
62%
60 % 28
PSNR
40 %
24
20 %
0% 20
0 5 10 15 0 5 10 15
Number of multiplexed images Number of multiplexed images
(a) (b)
41
31
32.5 dB
31
PSNR
28.5 dB
26
21
16
0% 20% 40% 60% 80% 100%
Compression rate
(c)
Fig. 2. Compression rate as a function of the number of multiplexed images without

optimization. (b) As in (a) for the PSNR. (c) A comparison of the PSNR values versus the
compression rate with our optimized method (squares) and the JPEG method (circles). The
circled values indicate that our scheme may not have better compression performances than
JPEG in the high PSNR case. The dashed line corresponds to the case of five multiplexed
images. The solid lines are guides for the eye.
Visual observation indicates that the quality of images is good and is better than the one
obtained with JPEG as shown in the bottom lines of Table 6 and Table 7. However, our
calculations show that our method may not have better compression performances than JPEG
in the high PSNR case.
Table 7. A comparison of the compression performance between JPEG (left) and SFCE
(right) methods
Tc JPEG Reconstructed image Tc SFCE Reconstructed image
7% 8%
PSNR= 36.2 PSNR=33.1
50% 52%
PSNR=31.4 PSNR=33
85% 86%
PSNR=28 PSNR=28.7
93% 94%
PSNR=25.5 PSNR=27.6
4. Optimized encryption adapted to the SFCE method

The spectral plane resulting from the fusion of the different DCT blocks constitutes a first
encryption level (Fig. 1). Without knowledge of the size of each block, its rotation, shape, and
position, it remains difficult to reconstruct the target images. Additionally, if DRP is applied
to the JPEG implementation, compression rate is limited, i.e. Tc ≤ 47% , since the encrypted
images are complex valued. Hence, a second level of encryption will be added without
degradation either in reconstructed image quality or in compression rate.
4.1 Optimization using real key image
Figure 3 shows the basic principle used to add a second encryption level. For that we used a
real key image (a single personal fingerprint). The second encryption key is found offline. We
start by Fourier transforming the biometric image, i.e. fingerprint, multiplied by a first random
phase mask. Next, only the real part of the spectrum is selected. Using the encrypted real-part,
the maximum value (max) is calculated. Then, a random mask with real values within the
range [0, max] is performed. Then, the real part of the encrypted spectrum of this fingerprint
is added to the second mask. It is interesting to notice that introduction of this random real
mask allows us to smoothen the values in the spectral multiplexed plane. Finally, the random
amplitude values are used to encrypt the spectral plane containing the different multiplexed
spectra. Thus, compressed and encrypted output plane is obtained.
N
Filter (1)
DCT
I(1) X I(1)
First 4 filtered Compressed and

Shift (1) DCT encrypted images :
DCT Real values
I(2) X I(1)
Shift (2) ...
Filter (2)
...
...
.
.
.
Shift (Ni)
DCT
I(Ni) X I(1)
Filter (Ni)
Spectrum
Key image Spectrum real part
FT
+
maximum value
x
« max »
Random phase Random values

key (1) [0 ; max]
key (2)
Fig. 3. Synoptic diagram illustrating the optimized encryption method.
We start by fabricating the decryption mask. This requires the knowledge of the key image
(fingerprint) and the procedure. The reverse procedure which is illustrated in Fig. 1 is then
performed. Here, the encryption rate is calculated via the PSNR between the target image and
the decrypted image without knowledge of keys. The PSNR should be larger than 15 dB for
reconstruction integrity. The technique achieves PSNR< 4 up to 16 images indicating a good
performance of encryption (PSNR<<30 dB i.e. the value of reconstructed image without
compression). Table 8 shows the PSNR corresponding to the encryption results as a function
of the number of multiplexed input images. Our calculations indicate that PSNR< 4dB.
Table 8. Encryption using a real key image: PSNR as function of the number of
multiplexed images
Number of multiplexed target

images 2 4 6 9 16
Calculated average of PSNR 3.32 3.47 3.57 3.64 3.60
values over all target images
For comparison, we also show calculations of Tc with encryption (Table 9) and m=5. The
performances are good since Tc is decreased by only 8% compared to the case where no
encryption is performed.
Table 9. Compression rate with m = 5
Number of multiplexed
Tc (%)
target images
2 13.5
4 56.8
6 71.3
9 80.8
16 89.6
4.2 Resistance against attacks

We confine our attention in this subsection to quantify the performance of our algorithm
against several plaintext-chosen attacks. Attack 1 is the most probable since the cipher does
not have any information about the target images, the used technique, and the encryption keys.
Attack 1: Our first test is conducted with the cipher having no information about our
technique. We assume that he or she has no knowledge of the number of target images, their
rotation and quantization, the encryption keys, and the encryption protocol. To accomplish his
or her goal, the cipher needs all this information which renders his or her task arduous.
Indeed, within the optimized SFCE algorithm that is presented here one needs to have the
fingerprint, random phase mask, quantization matrix, block sizes, their rotations and shapes to
decrypt the input information. The size of the random matrix and fingerprint is t x × t y . To
show how it is difficult to retrieve this information contained in a quantization matrix whose
size is equal to t x N i let us consider an example. We now briefly discuss how the combination
×
number scales
×
for the case of an image of size 256 × 256 : 256256 256 for the fingerprint and
(1015)256 256 for the random matrix. The quantization matrix (for an image of size 256 × 256)
has size 256 × 1. The latter includes maximum values of the DCT of the image. The
simulation of this matrix seems to be unlikely. In this case, testing all keys is discouraging.
Attack 2: The second test considers that the cipher now knows the principle of the optimized
SFCE. For simplicity, a single image was used in this test. Since the cipher has access to our
system, he or she can send an image corresponding to the inverse DCT (IDTC) of a matrix
whole elements are all equal to 1. Matrix quantization is realized such that each value is m-bit
encoded, i.e. every value is multiplied by (2m-1 – 1). Then, encryption block is realized. This
block is the sum of the real part of the Fourier transform of the fingerprint and of the
maximum value of the latter multiplied by a random matrix. The cipher can guess the
encryption block in suppressing a matrix of the same dimension as the image sent but for
which all the values are equal to 2 m−1 − 1 in the output plane of SFCE resulting from
encryption. Thus, for another sent image, he or she already knows the block which allows
encryption of this image. He or she suppresses it from the encrypted image and since the
quantization matrix is unknown to him or her, only the IDCT needs to be carried out. In Fig.
4, we show an example of what can be obtained by the cipher following his or her attack. As
displayed in Fig. 4(c), the resulting reconstructed image differs from the target image.
However, the latter shows some part of the information contained in the former.
(a) (b) (c)

Fig. 4. Example of attack with knowledge of the fingerprint and random matrix: (a) target
image, (b) compressed and encrypted image following the SFCE algorithm, (c) the cipher
result.
Optimization against attack 2: To get better results and enhance the robustness of our system
against such attack (2), one elegant idea for addressing the flaw is to divide the key image into
several parts, see, e.g. Fig. 5(a) and Fig. 5(b). Then, we introduce the number of the part sent
in the encryption process in such a way that the encryption block changes at each sending.
Consequently, even if the cipher knows the encryption block following this method, he or she
could not decrypt images sent later. The solution to this defect consists in changing the
numbered parts (Fig. 5(b)) according to the considered image. This implies a change of the
Fourier transform and of the encryption block with which the image is encrypted.
Interestingly, the number of the part can be easily introduced in the fingerprint. The output
encrypted image resembles to that shown in Fig. 4(b). However, only low values of PSNR
were obtained, e.g. after 5000 trials and using two of the four parts of the key image (Fig.
5(b)), PSNR ≤ 7.3 dB.
(a) (b)
Fig. 5. (a) Key image, (b) key image decomposed in several parts which have been permuted.
Thirdly, we assume that the cipher knows that a specific image exists among those to be sent.
He or she knows also that the second key is a fingerprint. If he or she finds a satisfactory
PSNR between the result he or she recovers the image to be sent (PSNR>18 dB), then the
algorithm stops. Otherwise, the cipher changes the value of the fingerprint considered using 4
× 4 blocks, then using 2 × 2 blocks, and finally pixel by pixel. At each change, the PSNR is
calculated and the algorithm stops if the PSNR is larger than 18 dB. The result of this attack
has not permitted recovery of the image.
Attack 3: Finally, we consider that the cipher knows a part of the image to be sent (part
( P × P ) of the target image depicted in Fig. 6) and where it is located. Once again, we assume
that the cipher knows the principle of the encryption method. So, he or she tries to decrypt the
image thanks to a given fingerprint. The algorithm stops when PSNR>18 dB. The PSNR is
calculated between the image reconstructed by the attack for which the image part and its
position are known. It is important to point out that we were unable to decrypt the target
image after trials lasting more than three weeks (P=64 pixels, image size=256 pixels). In
conclusion, this set of tests demonstrates the robustness of our algorithm against such kind of
attacks since even with knowledge of much data the original target image cannot be
decrypted.
Fig. 6. For the attack 3, the part of the image framed in red is known by the cipher.
5. Summary and conclusions

In conclusion, we have investigated several ways of optimization of the SFCE method
originally suggested in [13]. The presented technique has great potential to achieve
compression and encryption of images simultaneously in a dependent way. The motivation for
much of this research was the premise that control of the size, rotation, and shifting of DCT-
blocks can be achieved to optimize the bandwidth of the spectral plane. Finally, we stress that
the results presented here provide strong evidence that the use of an encryption method based
on real keys allows us to increase the robustness of this method against several types of attack
while maintaining a satisfactory compression rate. These results will hopefully stimulate more
theoretical efforts to obtain a greater understanding of efficient simultaneous compression and
encryption of multiple images, which, in turn, could stimulate the optical implementation of
these procedures.
Acknowledgment
We thank the reviewers of this manuscript for enhancing the quality of our work. This work
was supported by the Lab-STICC which is Unité Mixte de Recherche CNRS 6285. This
paper is dedicated to the memory of Alfalou’s mother, Mrs. Fatma Bayda.
View publication stats

Citation 16 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Citation 16 PDF

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Assessing the performance of a method of simultaneous compression and

Article in Optics Express · April 2013

Ayman Alfalou C. Brosseau

SEE PROFILE SEE PROFILE

DSP power efficiency View project

Magnetic materials View project

The user has requested enhancement of the downloaded file.

Abstract: We introduce a double optimization procedure for spectrally

References and links

(b) (c) (d)

size of multiplexed DCT spectral plane N2 1

Bloc size Compression rate MSE Reconstructed Number of

3.3. Calculation of the compression rate without quantization

Bloc size Compression rate (pixels) Compression rate

PSNR = 28.7 PSNR = 27.1 PSNR = 23.3

3.5. Quantization by block optimization and normalization between 0 and max

Number of PSNR Reconstructed image Reconstructed image

Fig. 2. Compression rate as a function of the number of multiplexed images without

Tc JPEG Reconstructed image Tc SFCE Reconstructed image

PSNR= 36.2 PSNR=33.1

4. Optimized encryption adapted to the SFCE method

First 4 filtered Compressed and

Random phase Random values

Fig. 3. Synoptic diagram illustrating the optimized encryption method.

Number of multiplexed target

4.2 Resistance against attacks

(a) (b) (c)

5. Summary and conclusions

View publication stats

You might also like