[Downloaded from www.aece.ro on Wednesday, July 02, 2014 at 07:39:33 (UTC) by 182.71.20.37. Redistribution subject to AECE license or copyright.]
Advances in Electrical and Computer Engineering
Volume 13, Number 3, 2013
Enhanced Segment Compression Steganographic Algorithm Ioan Gabriel BUCUR, Daniela STANESCU, Mircea STRATULAT University Politehnica of Timisoara, 300006, Romania
[email protected] 1
Abstract—Steganography is the science and art of concealing messages using techniques that allow only the sender and receiver to know of the message’s existence and be able to decipher it. In this article, we would like to present a new steganographic technique for concealing digital images: the Enhanced Segment Compression Steganographic Algorithm (ESCSA). We start by mentioning several desired properties that we have taken into consideration for our algorithm. Next, we define some quality metrics with which we can measure how well / to what extent those properties are achieved. A detailed presentation of the component parts of the algorithm follows, accompanied by quantitative analyses of parameters of interest. Finally, we discuss the strengths and weaknesses of our algorithm. In addition, we make a few suggestions regarding possible further refinements of the ESCSA. Index Terms—digital image, digital signatures, image coding, information security.
I. INTRODUCTION In steganographic communication, two aspects are especially important. First, the messages delivered to the receiver by the sender have to be imperceptible to the eye of possible attackers. Only in this situation can the communication truly be secret. Second, the embedding capacity has to be as large as possible, so that the maximum amount of information can be hidden [1-2]. The Enhanced Segment Compression Steganographic Algorithm (ESCSA) addresses these two necessities by combining the compression power of the Karhunen-Loève Transform (KLT) with the embedding capacity of the Least Significant Bit (LSB) substitution. The imperceptibility and the embedding capacity are two parameters at odds with each other. In order to increase the imperceptibility of the steganographied image, the embedding capacity must be limited or even reduced. Conversely, in order to increase the embedding capacity, the image’s imperceptibility must be sacrificed to a certain degree. One way to ameliorate one parameter without having a negative effect on the other is compression. The Karhunen-Loève Transform, also known as the Hotelling Transform or Eigenvector Transform allows an optimal compression, superior for instance to the one achieved by the popular Discrete Cosine Transform (DCT), the latter being in fact just an approximation of the KLT [3]. The KLT completely decorrelates the input signals and is able to reallocate their energy in just a few components. The best KLT numerical algorithms usually involve heavy data processing that demands significant
computational resources and produces lengthy execution times. On the other hand, the DCT numerical algorithms are much more computationally efficient and have a much lower execution time, which is why the DCT has been used more frequently for compression purposes in steganographic algorithms until now. We plan to overcome this disadvantage of the KLT by means of segmentation (dividing the digital image into segments), so that we can benefit from its higher compressing power. II. STEGANOGRAPHIC PROPERTIES AND QUALITY METRICS In order to evaluate the performance of our Enhanced Segment Compression Steganographic Algorithm, we need to define a set of steganographic properties that express the quality of a given steganographic algorithm. To be able to compare the ESCSA with other steganographic algorithms, we also have to define one or more quality metrics for each steganographic property. Through these quality metrics, we can express the steganographic properties of an algorithm in a quantitative way. A.
Embedding Capacity
The embedding capacity represents the maximum number of bits that can be embedded in the cover image when using a certain steganographic system / algorithm [4]. If we consider a simple algorithm that uses the LSB substitution for embedding, we can easily calculate the embedding capacity to be b/8 bits for a cover of b bits. This is because only the least significant bit from the eight bits of a byte is used for embedding. The embedding capacity is closely related to the steganographic capacity: the maximum number of bits that can be embedded in a cover image so that the hidden payload remains imperceptible. A metric that allows us to express the embedding capacity quantitatively is the compression ratio, also known as the compression rate [5]. This metric is calculated as the ratio between the size of the compressed payload (after applying the Karhunen-Loève transform on it) and the initial size of the payload. B. Imperceptibility The imperceptibility of a steganographic algorithm is a rather subjective property. It indicates the capacity of a steganographic technique to embed payloads so that they are invisible to the attackers (“hidden in plain sight”).
1
Digital Object Identifier 10.4316/AECE.2013.03016
101 1582-7445 © 2013 AECE
[Downloaded from www.aece.ro on Wednesday, July 02, 2014 at 07:39:33 (UTC) by 182.71.20.37. Redistribution subject to AECE license or copyright.]
Advances in Electrical and Computer Engineering
AD
W
1 3HW
H
( SRxy CRxy
x 0 y 0
1 MSE 3HW
W
H
(1)
SGxy CGxy SBxy CBxy )
SRxy CRxy SGxy CGxy SBxy CBxy 2
2
x 0 y 0
2
(2) (3)
RMSE MSE
PSNR 10log10
Volume 13, Number 3, 2013
2
MAX MAX 20log10 MSE RMSE
(4)
H ,W image height and width SRxy , SGxy , SBxy color component in stegoimage CRxy , CGxy , CBxy color component in normal image
To measure the imperceptibility property, we will use the level of cover distortion, which we will often call the coverstego error. The cover-stego error can be expressed through different metrics: the root-mean-squared error (RMSE) [4],[6], the average absolute difference (AD) [7] or the peak signal-to-noise ratio (PSNR) [6-7]. The mathematical formulas of these metrics are given in (1)-(4).
III. ESCSA – STEP BY STEP DESCRIPTION The Enhanced Segment Compression Steganographic Algorithm, like any other steganographic algorithm, is composed of two perfectly mirrored parts: obtaining the steganographied image (stego), which takes place at the sender level, and recovering the payload, which takes place at the receiver level. A. Segmentation We divide the payload horizontally, in such a way that the segments are in fact clipped horizontal strips. The segmentation is described by the parameter segm_size, which suggestively indicates the size of the segments. If the payload has dimensions height x width, then the segments will be segm_size x width image blocks. In the next step, the compression will be performed on each segment separately and not on the whole cover image at once. Segmentation solves the computational complexity problem of the Karhunen-Loève transform, thus allowing a
fast and at the same time powerful compression. This kind of compression will permit us to double the algorithm’s embedding capacity in most practical cases. In Fig.2, we present the execution time of the ESCSA with and without segmentation enabled. We define the metric speedup as the ratio between the execution time of the algorithm with segmentation disabled and the execution time of the algorithm with segmentation enabled. The tests in Graph 1 were performed on an Intel i5-670 quad-core 2.8 GHz processor with 4GB DDR3 SDRAM, without other user applications running in the background. When comparing the execution times in Fig. 2, we can see that segmentation does not yield spectacular results in terms of payload recovery speedup. We only have a speedup value of around 5-10. In this case, the speedup results mainly because in the segmented versions the multiplications needed for payload recovery are performed on smaller matrices. This will result in a smaller number of elementary multiplication operations, which in turn leads to a shorter execution time. When we consider the speedup in generating the stego image provided by segmentation, the results are much more interesting, as can be seen in Fig. 2. This speedup has bigger values for larger payloads. Concretely, we obtain a speedup of around 200 for images of dimensions 200x135 (“watch” and “flowers”). This speedup increases to around 2500 for a payload of dimensions 640x480(“dogs”). For concealing the payload “dogs”, we managed to decrease the execution time through segmentation from 158 seconds to only 60 milliseconds. We conclude that segmentation makes the application of the KLT compression practical, permitting reasonable execution times. An interesting observation is that the speedup is almost independent with regard to the dimensions of the original cover image (and by extension to those of the stego image). Indeed, in graph 1 we see very similar speedups for cover images of different sizes, when the payload dimensions are maintained constant. For example, this is the case for the covers “lena” and “building”, when they conceal the payload “watch”. This can be easily explained by the fact that the cover image is a passive object with respect to the
Figure 1. Enhanced segment steganographic compression process
102
[Downloaded from www.aece.ro on Wednesday, July 02, 2014 at 07:39:33 (UTC) by 182.71.20.37. Redistribution subject to AECE license or copyright.]
Advances in Electrical and Computer Engineering
Volume 13, Number 3, 2013
algorithm. Its only role is to be the carrier to the secret message, so it is only involved in the embedding and extracting phases. The payload is the active object, on which almost all the processing is done. It is thus natural that its characteristics (dimensions, colour composition) will have a huge influence on the execution times and potential speedups. B. Compression We will not enter into the details of the KLT algorithm [3], [8-9]. Instead, we will show, by means of a graph, the significant compression ratio that can be achieved through this method. We used various images for testing. This allowed us to observe that the KLT compression ratio varies widely depending on the colour composition of each payload. In Fig.3 we can observe compression ratios ranging from 15% to 71%.
building‐dogs
building‐album
lena‐numbers
lena‐mermaid
building‐mermaid
building‐watch
lena‐watch
3000.00 2500.00 2000.00 1500.00 1000.00 500.00 0.00 lena‐flowers
speedup
Stego image generation speedup
cover‐payload
Payload recovery speedup
Figure 3. Compression ratios
The minimum value implies that a particular image was compressed 6.66 times, meaning it requires a memory storage space 6.66 times smaller. On average, we obtain a compression ratio of around 33%, which means that we can expect to compress a new payload around three times. C. Embedding For embedding purposes, we use the technique of hiding the LSBs of the cover image’s bytes. LSB substitution simultaneously ensures a large embedding capacity and a good visual imperceptibility [1],[10-12]. These properties are confirmed by the small cover-stego error obtained in test results[13-15] . In Fig. 4 we show the cover-stego errors for various cover-payload pairs, as described by the NRMSE metric. The NRMSE (normalized RMSE) is simply the RMSE metric that we defined earlier divided by the maximum value of a RGB pixel’s component. We use normalization to change the range values of the RMSE from [0, MAX] to [0,1]. This allows us to express the cover-stego error in a more suggestive way, namely percentages. According to Fig. 4, the obtained NRMSE is at most 1%, all cases considered. This value suggests a small cover-stego error and consequently a high imperceptibility of the stego image.
12.00
speedup
10.00 8.00 6.00 4.00 2.00 0.00
cover‐payload Figure 2. Segmentation based stereo image generation and payload recovery speedups
Figure 4. Cover-Stego error
103
[Downloaded from www.aece.ro on Wednesday, July 02, 2014 at 07:39:33 (UTC) by 182.71.20.37. Redistribution subject to AECE license or copyright.]
Advances in Electrical and Computer Engineering
Figure 5. Stereo image and payload recovery time
D. Parallelization An aspect on which we have focused particularly when building the ESCSA is the execution time. When the execution time of a steganographic algorithm is too lengthy, we are constrained to eliminate key processing operations, thus stripping the algorithm to a bare minimum (embedding and extraction). On the other hand, if we manage a short execution time, we gain the possibility to add supplementary features meant to improve on the other desired steganographic properties. Consequently, we deem the execution time to be of utmost importance in the development of a steganographic algorithm.
104
Volume 13, Number 3, 2013
Figure 6. Parallelization based stereo image generation and recovery speedup
Consistent with our philosophy, we have tried to reduce the execution time of the ESCSA whenever possible, through various means of optimization. Such a means is execution parallelization. Due to segmentation, the ESCSA presents a good amount of thread-level parallelism. Every segment of the payload can pass through all the processing stages (compression, quantization and finally embedding) almost independently of all the other segments. Indeed, there are only few execution points where the segments’ processing needs to be synchronized. In Fig. 5 we can observe how the ESCSA runs in a multithreaded environment. We performed multiple time measurements for each cover-payload pair and calculated their mean. The mean execution times are those displayed in Fig. 5. To get an even better overview of the benefits
[Downloaded from www.aece.ro on Wednesday, July 02, 2014 at 07:39:33 (UTC) by 182.71.20.37. Redistribution subject to AECE license or copyright.]
Advances in Electrical and Computer Engineering achieved through parallelization, we also present the parallelization speedups in Fig. 6. We note that the multithreaded measurements were performed using four separate execution threads, each assigned to a distinct core of our quad-core processor. As expected, the improvement in execution time resulting from parallelization is significant. In all tests, the ESCSA performs very well, not only in the multi-threaded environment, but also in the single-threaded one. All the execution times measured are lower than half a second, a decent performance for most practical steganographic application. Moreover, Fig.6 shows that as the payload gets progressively larger (from “watch” and “flowers” of sizes 200x135 to “football” of sizes 1600x1200), the speedup is greater, tending to the maximum theoretical limit of four. This phenomenon is easily understandable since more processing time is needed for larger images, while the overhead (thread management) remains constant, thus constituting a smaller portion of the total execution time. IV. CONCLUSION The strong features of the Enhanced Segment Compression Steganographic Algorithm place it in a good spot for practical applications. As mentioned, the algorithm is designed and optimised for concealing digital images in other digital images. The ESCSA’s greatest strengths are the excellent embedding capacity provided by the KLT compression and the good visual imperceptibility provided by the LSB embedding technique[16-18]. It is also very important to restate that the ESCSA has a very short execution time, given the computational complexity of image processing [19-20]. The algorithm’s inherent concurrent nature recommends it for deployment on multicore platforms, for instance intelligent mobile devices with image processing capabilities. The ESCSA performs less well in terms of robustness and security. However, we have taken steps and measures to improve the metrics of these properties. For instance, we have incorporated Hamming error detection and correction codes in our algorithm to improve robustness. In order to ameliorate security, we have changed the method of transmitting the encoding parameters to the decoder (the socalled Composite Key – see algorithm flowchart) from unencrypted to private key encrypted. Further refinement ideas include randomising the embedding zones in the cover image in order to increase robustness or encrypting the payload before embedding to ensure better security. Furthermore, we will investigate the implementation of this algorithm on FPGAs, which have recently have been demonstrated to have good performance for division and square root operations [21], on one hand, and on modern mobile devices, which present a very good performance/power consumption ratio. The experimental results are based on the applying the proposed algorithm on 16 color and grayscale images depicted in Fig. 7-14, which have been used from [22-37]. These images have various sizes (from 200x135 pixels up to 4877x3515 pixels).
Volume 13, Number 3, 2013 APPENDIX A – IMAGES
USED IN VALIDATION PROCESS
a) b) Figure 7. “sparrow” (800x600)[22] (a) and “hawk”(800x600) [23] (b)
a) b) Figure 8. “mermaid” (256x256)[24] (a) and “lena”(512x512) [25] (b)
a) b) Figure 9. “sphinx” (800x600)[26] (a) and “winter”(600x400) [27] (b)
a) b) Figure 10. “numbers” (256x256)[28] (a) and “peppers”(600x400) [29] (b)
a) b) Figure 11. “flowers” (200x135)[30] (a) and “watch”(200x135) [31] (b)
a) b) Figure 12. “star” (4877x3515)[32] (a) and “football”(1600x1200) [33] (b)
105
[Downloaded from www.aece.ro on Wednesday, July 02, 2014 at 07:39:33 (UTC) by 182.71.20.37. Redistribution subject to AECE license or copyright.]
Advances in Electrical and Computer Engineering
Volume 13, Number 3, 2013
[12]
[13]
a) b) Figure 13. “building” (640x480)[34] (a) and “dogs”(640x480) [35] (b)
[14]
[15] [16]
[17] [18]
a) b) Figure 14. “car” (1024x1036)[36] (a) and “album”(400x400) [37] (b)
[19] [20]
REFERENCES [1]
Frank Y. Shih, Digital Watermarking and Steganography: Fundamentals and Techniques, CRC Press, ISBN 978-1-4200-47578, Boca Raton, FL, SUA, 2008 [2] Hajizadeh, m., Helfroush, m. s., Dehghani, m. j., Tashk, a., A Robust Blind Image Watermarking Method Using Local Maximum Amplitude Wavelet Coefficient Quantization, Advances in Electrical and Computer Engineering, vol.10, Issue 3,ISSN: 1582-7445,2010 [3] S.G.Hoggar, “Mathematics of Digital Images”, Cambridge University Press, 2006, ISBN-13 9780521780292 [4] Ingemar J. Cox et al, Digital Watermarking and Steganography, 2nd edition, Morgan Kaufmann, ISBN 978-0-12-372585-1, Burlington, MA, SUA, 2008 [5] Daniela Stănescu, Ioan-Gabriel Bucur, Mircea Stratulat, Segment Compression Steganographic Algorithm, International Joint Conference on Computational Cybernetics and Technical Informatics (ICCC-CONTI), ISBN 978-1-4244-7432-5, 2010, pp. 349-354 [6] Eric Cole, Hiding in Plain Sight: Steganography and the Art of Covert Communication, Wiley Publishing, ISBN 0-471-44449-9, Indianapolis, IN, SUA, 2003 [7] Stefan Katzenbeisser, Fabien A.P. Petitcolas, Information Hiding Techniques for Steganography and Digital Watermarking, Artech House, ISBN 1-58053-035-4, Norwood, MA, SUA, 2000 [8] Dafas P., Stathaki T., Digital Image Watermarking using block-based Karhunen-Loeve Transform, Image and Signal Processing and Analysis, ISPA2003, Poceeding of the 3th International Symposium, 18-20, sept. 2003, ISBN: 953-184-061-X, pp1072-1075, vol.2 [9] Michael Gastpar, Pier Luigi Dragotti, Martin Vetterli, The Distribute Karhunen-Loeve Transform , IEEE Transaction on Information Theory, vol. 52, no.12, dec. 2006, pp 512-522 [10] Gregory Kipper, Investigator’s Guide to Steganography, CRC Press, ISBN 0849324335, Boca Raton, FL, SUA, 2004 [11] Daniela Sănescu, Valentin Stângaciu, Ioana Gergulescu, Mircea Stratulat, Steganography on Embedded Device, 5th International
106
[21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38]
Symposium on Applied Computational Intelligence and Informatics, SACI 2009, ISBN:978-1-4244-4478-6, Timisoara, 2009, pp.313-317 Boncelet C., MarvelL., Lossless Compression- Based Stegnalysis of LSB Embedded Images, 41st Annual Conference on Information Sciences and Systems, CISS’07,Baltimore, 14-16 March 2007, ISBN:1-4244-1037-1, pp.923-926 Feng H., Effros M., On the rate-distortion performance and computational efficiency of the Karhunen-Loeve Transformfor lossy data compression, IEEE Transaction on Image Processing, feb. 2002, vol.11, Issue2, pp.113-122 Ki-Hyun Jung, Kyeoung-Ju Ha, Kee- Young Yoo, Image Data Hiding Method Based on Multi-Pixel Differencing and LSB Substitution Methods, International Conference on Convergence and Hybrid Informatin Technology, ICHIT”08, Daejeon, 28-30 Aug., 2008, ISBN: 978-0-7695-3328-5, pp. 355-358 Xiaolong Li, Tieyong Zeng, Bin Yang, Improvement of the Embedding Efficiency of LSB Matching by Sum and Difference Converting Set, IEEE International Conference on Multimedia, Hannover, June 23, 2008, pp.09-212 W. Burger, M. Burger, Digital Image Processing, Springer, 2008, ISBN:978-1-84628-379-6 Chang C-C., Chou H., Lin c-C., Colour Image-hiding scheme using human visual system, Imaging Science Journal, Oxford, UK, sept. 2006, vol. 54, nr.3, pp.152-163 Eric Cole, Hiding in Plain Sight: Steganography and the Art of Convert Communicating, Wiley Publishing, Inc., Indianapolis, SUA, ISBN: 0-471-44449-9, 2009 Kawaguci E , Eason R., Large Capacity Steganography, U.S.Patent no. 6,473,516, oct. 29, 2002 He Junhui Tang, Shaohua Wu Tingting, On the Security of Steganographic Techniques, Congress on Image and Signal Processing, CIPS, 2008, China, 27-30, May, pp716-719, vol.5 A. Amaricai, O. Boncalo “Implementation of very high radix division in FPGAs” IET Electronics Letters, Vol.48, Issue 18, 2012 http://www.wallbank.me.uk/gallery/albums/userpics/sparrow_001_6 00x400.jpg http://lindsaywenzel.com/ColeByrd.com/Images/battle%20hymn/eag le %25205.jpg http://www.freewebs.com/riccosite/Aquaria_256x256.png http://1.bp.blogspot.com/_6CKi9y4Mf88/SPHkHSnRDI/AAAAAAA AAAY/mPaBsHTg3JE/s1600/lena760142.bmp http://www.paragontravel.com/wp-content/uploads/2010/08/EgyptThe-Great-Sphi5DAAE0.jpg http://i579.photobucket.com/albums/ss232/myerycika/90_09_7--Winter-Scene--Northumberl.jpg http://hypermultimedia.com/x3d/URLTests/256x256.jpg http://www.stanford.edu/class/ee368b/Projects/cnetzer/peppers512x 512.jpg http://scien.stanford.edu/pages/labsite/2005/psych221/projects/05/vvi k ram/wildflowers.jpg http://whiterockhealthcenter.com/images/img-06.jpg http://www.redorbit.com/media/gallery/universe/iod-universe042712.jpg http://www.kostenlosewallpaper.com/wallpapers/fussball-hdwallpaper-kostenlos--1600x1200.jpg http://www.nsuok.edu/goodies/wallpaper/brokenarrow640x480.jpg http://anatrofin.files.wordpress.com/2009/05/33-cute-puppies.jpg http://newedgemustangs.com/site%20images/mach_red2_1024.jpg http://www.mp3boo.com/cover-album/abbey_road.jpg
