INTRODUCTION
Gone are the days when images were only about memories of the past. The images
now speak more than that because of the advent of the field of image steganography
(Cheddad et al., 2010), which embeds secret information
in images imperceptible to the naked human eye (Rabah, 2004;
Amirtharajan and Balaguru, 2009). From time immemorial
emphasis on new techniques for secret communication has been given high importance
based on the levels of confidentiality required. Three interlinked techniques
namely cryptography (Schneier, 2007), steganography and
watermarking (Stefan and Fabin, 2000) form the base for
secure communications. While, cryptography involves making the content undecipherable
the other two are information hiding methods where the mere presence of information
is itself hidden (Amirtharajan et al., 2011).
Steganography is an art of embedding the secret information within some other
file generally known as cover. The main objective of steganography is to provide
a secret communication between any users such that any undesirable person cannot
gain access to the information by just glancing at the cover file (Bender
et al., 1996). This prowess of secret communication coined as Steganography
means “covered writing” in Greek. Various methods are implemented
to obliterate the existence of the secret message. A bunch of such techniques
include invisible ink, character arrangement, microdots and digital signatures
and spread spectrum.
Steganography basically administers four main modules say, a cover file which
acts as the container for the secret message, a secret message that contains
the confidential data, a key for encoding the secret message and steganography
algorithm for sneaking in the secret message inside the container as depicted
in Fig. 1. The combination of these modules creates the stego-object
which is then sent to the receiver. At the receiver’s end the decoding
algorithm is implemented to retrieve the original message from the stego-file
(Wang and Wang, 2004).
In the recent times, digital media such as text, audio, video, protocols or
image are being opted for steganography (Rabah, 2004;
Bender et al., 1996). The sender decides the effectuation
based on their requirement. However due to the confined ability of the human
visual system, image steganography is best-loved among the available methods.
This is due to the presence of superfluous information in the images which can
be easily interpolated. Table 1 present differences among
steganography, watermarking and cryptography.
|
| Fig. 1: |
Simple schematic of stego system |
| Table 1: |
Different among steganography, watermarking and cryptography |
 |
| Table 2: |
Year wise publication details |
 |
| Table 3: |
Total number of publication in web search |
 |
| Table 4: |
Total number of patents filed in various patent offices |
 |
|
While the ambit of steganography is immense, it has been mainly used for secret
communication. Although, steganography and cryptography both cater to the same
purpose, the advantage of former over later is that, it obliterates the existence
of the secret message (Thanikaiselvan et al., 2011a;
Padmaa et al., 2011; Janakiraman
et al., 2012). It can be used to hide a systems identity by masking
the TCP/IP headers. Apart from these many other task like feature tagging such
as time stamp, captions etc can be embedded in an image. Steganography has also
found its way in various technologies such as X-ray (containing both doctor’s
notes as well as the X-ray image), printers (scarcely visible tiny yellow dots
indicating the printer serial number etc. (Petitcolas
et al., 1999; Zaidan et al., 2011).
In order to enhance the security of the communication, the information is first
encrypted and later hidden so that the attacker has to first find the blotted
out information and then decrypt it (Padmaa et al.,
2011). The flaw with cryptography alone is that the encrypted message can
be easily ascertained during transit by anyone who is looking for it thus making
the communication vulnerable. While aspiring for better outcome in image steganography
it is important to keep three aspects in mind. The first one is the capacity
of the cover image i.e., it should be able to store utmost data inside it. Second
being imperceptibility should provide an unaltered image after data hiding (Zaidan
et al., 2011). Last but not the least is robustness i.e. it should
be secure against attacks. Blending these three aspects efficaciously will help
in achieving the goal of image steganography.
To justify the popularity of the chosen review topic, a simple search has been
conducted, to survey the number of articles, which have been published in referred
and peer reviewed journals and number of Patents filed in various patent offices
throughout the world with the search key word Steganography and the results
are presented in Table 2-4.
Search key word: Streganography.
Web search sources: Science Direct, SCOPUS, IEEE, Springer, DOAJ, Citeseer, Sciverse, ACM Portal and Google Scholar.
Table 2-4 confirm that, year by year the
number of papers published in the chosen field of study has dramatically increased.
Furthermore, there are three different review/survey papers available in the
existing literature for steganography (Petitcolas et
al., 1999; Li et al., 2011; Chen
and Shi, 2008; Cheddad et al., 2010), but
none of the three specify anything about random image steganography. Hence,
in this study, the present status and future direction has been presented.
STEGANOGRAPHY BASIC TERMINOLOGY
| • |
Cover image is the carrier which is used to carry or preserve
the embedded bits |
| • |
The image embedded with the secret data is referred to as the stego-image |
| • |
Capacity or simply Payload is amount of secret data to be embedded and
sometimes expressed in Bits Per Pixel (BPP) or simply in number of bits |
| • |
Imperceptibility - literal meaning is very slight or hard to find the
difference between the cover image and Stego Image |
| • |
Robustness-with standing ability even towards intentional attacks but
not suitable for Steganography, because if somebody suspects, there is a
covert communication, then steganography is no more undercover operation |
| • |
Steganalysis includes the attacks on the images by various image processing
techniques to expose the hidden information by attackers |
REVIEW ON THE STEGANOGRAPHY
Steganography, derived from Greek words meaning ‘covered writing’
has been in use over the past thousands of years (Stefan and
Fabin, 2000). For example, Histaiacus tattooed a message on a slave’s
shaved head and the slave was sent as messenger after his hair grew back. Another
way of secret writing of the Chinese was reinvented by the Italian mathematician
Jerome Cardan and included a paper mask with holes between both the sender and
receiver and secret message is written by keeping the paper mask on a blank
sheet. Later the blanks are filled to appear as innocuous text and this method
is called as Cardan Grille. The Nazis invented several steganographic methods
during the World War II using invisible ink and null ciphers and Microdots.
An example can be given as follows: Apparently neutral’s protest is thoroughly
discounted and ignored. Isman hard hit. Blockade issue affects pretext for embargo
on by-products, ejecting suets and vegetable oils.” is a message sent by
a Nazi spy, which when decoded using the second letters reveals the secret message
‘Pershing sails from NY June 1’. Another interesting case is wherein
Morse code was concealed in a drawing in the form of long grass and short grass
in the year 1945. From these ancient methods, steganography has made a giant
leap to the digital form as well due to the enormous improvement in computer
power, internet and advancements in digital signal processing. Steganography
is widely used in confidential communication purposes and a continuous evolution
is guaranteed due to its applications. Any advancement in science would certainly
have a disadvantage. In this case, the concern was about the usage of steganography
by terrorists for secret plots, which is also called as cyber planning or digital
menace (Stefan and Fabin, 2000). Steganography does not
exist for still images alone. Embedding hidden messages in video (Al-Frajat
et al., 2010) and audio files (Zhu et al.,
2011) is also possible which makes steganalysis even tougher. The general
schematic description of steganography with different cover (Video, Audio and
Image) are shown in Fig. 2.
Rabah (2004) provides a brief history of steganography
and discusses about how it is related to cryptography. In addition to that it
also expounds the purpose, strengths, weaknesses, advantages and disadvantages
of steganography. A study was conducted by Hmood et al.
(2010a) wherein strengths and weakness for possible multimedia covers for
steganography approaches were elucidated and methods to improve capacity and
security were discussed. Hmood et al. (2010b)
in their work has presented an overview of sophisticated information hiding
techniques which involve image file as a cover for carrying out the processes.
In a High-capacity Steganographic Scheme for 3D Point Cloud Models by Qi
et al. (2010), Self Similarity Position Matching (SSPM) procedure
was used presenting high-capacity spatial side-match steganography for 3D point
cloud model. Mohammad et al. (2011) in their
work proposed a transparent and high hiding capacity algorithm which effectively
uses Block Truncation Coding (BTC) by employing a two level (one-bit) nonparametric
quantizer and Human Visual System (HVS) masking characteristics ensuring high
visual quality of stego image.
Hong et al. (2009) proposed a method to improve
the stego image quality by using reversible contrast mapping data hiding scheme
that uses the variance feature of the cover image.
|
| Fig. 2: |
General schematic description of Steganography |
Luo et al. (2011) proposed a method that incorporates
secret sharing and data hiding technique for block truncation coding for image
compression. Under this method two quantization levels are hidden under in two
share images to increase the level of compression. Zeki
et al. (2011) have developed a digital watermarking model which can
find out the possibility to embed maximum amount of data in an image without
quality reduction. Zaidan et al. (2011) have
done a review on the impact of data privacy and confidentiality on developing
telemedicine applications in this paper. They have also done a short survey
on 130 participants to support their claim. Alanazi et
al. (2010) have proposed a handshake protocol depending on symmetric
and asymmetric cryptography for patient, medical center and doctor, medical
center for improving the privacy electronic medical record transmission.
Zanganeh and Ibrahim (2011) have proposed a substitution
based technique which involves embedding data into uncertain and higher LSB
layers which allows flexibility and also allows us to embed a large amount secret
message. Padmaa et al. (2011) have proposed a
method that combines steganography and cryptography in a unique way so as to
provide enhanced level of security. Here we get two different forma of the message
T1 and T2 using two distinct keys K1, K2. Amirtharajan and
Balaguru (2009) have proposed a three layer method for image steganography.
First layer involves pixel statistics conversion method. Second stage involves
embedding the result into a gray image using moore space filling curve and finally
the third stage consists of the gray image being embedded into another image
using Hilbert space filling curve to produce the final colour image. Amirtharajan
and Rayappan (2012) also explained a way to reduce the embedding effect
and to improve the stego image quality by adapting different traversing path
along with inverted pattern approach. Amirtharajan et
al. (2010) filed a patent using kolam traversing path to improve the
cryptic effect. Amirtharajan et al., (2011) proposes
a method using Pixel Indicator along with adaptive embedding and integrity check
to enhance cryptic effect. Furthermore, a detailed survey on effective way to
combine OFDM+CDMA for secure communication using information hiding is available
by Thenmozhi et al. (2011).
Al-Azawi and Fadhil (2010) proposed a text steganography
technique suitable for Arabic characters, hiding information by inserting extensions
character (kashida). Huffman coding was employed at its initial phase to compress
text into binary format. Xiang et al. (2011)
designed a novel steganography method that employed selecting a series of MCQ’s(Multiple
Choice Questions) that could generate the secret data. This technique when compared
against existing linguistic steganographic techniques, seemed superior. Luo
et al. (2008a) presented an algorithm based on the directed Hamiltonian
path selection in the complete digraph mapped from multi-blogs with same article.
This work was found to possess good imperceptibility and high security.
This review suggested the following, that a simple classification could be
methods based on spatial (Amirtharajan and Balaguru, 2011a;
Thanikaiselvan et al., 2011b) or transform domain
techniques (Provos and Honeyman, 2003; Thanikaiselvan
et al., 2011a) or there is a possibility to classify steganography
methods based on the covers i.e., Video Steganography, Text based Steganography
(Shirali-Shahreza and Shirali-Shahreza, 2008) Audio
Steganography and Image Steganography. The former method could be further classified
into substitution, transform, statistical, spread spectrum (Kumar
et al., 2011) and Amirtharajan and Balaguru (2011b)
and cover generation methods. Another classification on steganography is pure
steganography, secret key steganography and public key steganography.
In the belles-lettres of steganography, there are three standard protocols namely pure steganography, public key steganography and secret key steganography.
Pure steganography: This system includes the pure or unadulterated working principles of steganography, wherein the prior transfer of shared/ secret key. As in the shared or secret key transfer to the other participant/recipient is not mandatory, but still the purpose can be achieved.
In the common terms, the embedding process in this mode of steganography is coined and elaborated in terms of mapping in set theory, where E be the embedding process and C be the set of all cover images and M be the set of all possible images and is defined as E: CxM→C while the extraction process is denoted by D and is a mapping representing D: C→M, which depicts the extraction of secret message off the cover.
It is mandatory that |C|> = |M| and the embedding and extraction algorithm must be accessible to both transmitter and receiver, but to no one else.
Secret key steganography: As pure steganography relies completely on
secrecy, it cannot be always banked upon when the transfer is prone to loose
secrecy, although the transfer is between E (Embedding End) and D (Extraction
End). This is not secure as it violates kerckhoff’s principle (Provos
and Honeyman, 2003), which simply states, the method of transferring must
be secure even though none knows its encoding process).
With this Secret Key steganography can be implemented with three stratified objects (A cover image C, Secret Message M and a Secret key K). The sender chooses random cover image takes the secret message embeds it into the cover image deploying the secret key K along with it. And at the receiver end, if the secret key is known, the secret message can be extracted from the cover image. And as obvious, any person unaware of secret key has no scope of finding out the secret message.
Mathematically, with set notation we have, EK: CxMxK→C and DK: CxK→M (where EK-Embedding with secret key and DK-Extraction with Secret Key K).
Public key steganography: As the name suggests the encoding key is public
key, which is visible to the public database and everyone has access to it.
So, the embedding process is visible to everyone. Here, the true sender will
also use a secret key D, which is the apt key for decoding the secret message.
So, although, the public key is known to the third person, the message cannot
be decoded until and unless the third person knows the secret key D. All the
three are the classification on existing steganography.
Pseudorandom permutation steganography: A little too deceiving method for the sender, as this incorporates the complete utilization of the cover image, such that every bit is not left turned. This method of employment of every bit of the image along with the secret key and in a random fashion of encoding leaves the hacker with no hint of find the subsequent encoded bits hence making it extremely tedious to poach into the data.
This random fashion of encoding can be done with multiple keys( k1, k2, k3 and so on) or by the creation of multiple element indices from (J1………Jm).
RECENT TRENDS SUGGESTED IN RANDOM IMAGE STEGANOGRAPHY
Random image steganography, defined as an image steganography which offers cryptic effect. So far existing literature expects and suggest that well defined cryptography algorithm could be employed prior to embedding on the confidential information to offer cryptic effect. But random image steganography is to preprocess the data with well defined cryptographic algorithm then adapt any one of the possible ways to improve the complexity.
Possible ways of random image steganography:
Method 1: K bit embedding with a key 1 [0 0 0 0 1 1 0 1] in a pixel
Method 2: Key 2 [1 0 1 0 1 1 0 1] in a cover
Method 3: Using Fibonacci series 1, 2, 3, 5, 8, 13 …. If exceeds
use mod length of the Cover, problem collision attacks
Method 4: Variable bit embedding on the pixel (PVD)
Method 5: Encrypt the secret then embed
Method 6: Use different encoding instead of ASCII
Method 7: Distort the cover with known value then embed
Method 8: Compress the secret data then embed
Method 9: Compress, Encrypt and Embed
Method 10: Compress+Error control codes+Encrypt then embed
Method 11: Select the pixels by Generating Pseudo random number then
Embed
Method 12: Select pixels based on intensity values then variable bit
embedding
Method 13: Study the statistics of the cover without affecting embed.
Else Better Change the “ROUTE” for embedding
Method 1: K bit embedding with a key 1 [0 0 0 0 1 1 0 1] in a pixel: It will embed data into the pixels based on the four LSBs of the key.
| • |
If key = [0 0 0 0 1 1 0 1] then embedding should be in the
1st (2^0), 3rd (2^2) and 4th (2^3) LSBs |
| • |
If key = [0 0 0 0 0 1 0 1] then embedding should be in the 1st(2^0)
and 3rd(2^2) LSBs |
Merits: Its embedding capacity, MSE and PSNR depend on the key. Randomness is introduced in this method if the keys are other than [0 0 0 0 0 0 0 1], [0 0 0 0 0 0 1 1], [0 0 0 0 0 1 1 1], [0 0 0 0 1 1 1 1] as in these cases it will be normal LSB substitution with 1, 2, 3, 4 bits, respectively.
Complexity:
| • |
If the data is encrypted using DES it will introduce a complexity
of (2^64) |
| • |
Probability of embedding 0-bits in last four LSBs is ((4c3)/16) = (1/16) |
| • |
Probability of embedding 1-bit in last four LSBs is ((4c1)/16) = (4/16) |
| • |
Probability of embedding 2-bits in last four LSBs is ((4c2)/16) = (6/16) |
| • |
Probability of embedding 3-bits in last four LSBs is ((4c3)/16) = (4/16) |
| • |
Probability of embedding 4-bits in last four LSBs is ((4c4)/16) = (1/16) |
| • |
So the final complexity of embedding 1-bit is: (2^64)*(1/16)*(1+4+6+4) |
Complexity: Good when keys [0 0 0 0 0 0 0 1], [0 0 0 0 0 0 1 1], [0
0 0 0 0 1 1 1] and [0 0 0 0 1 1 1 1] are not used.
Imperceptibility: Visible to some extent if the key used is [0 0 0 0 1 1 1 1].
Suggestion: Don’t embed more than 3 bits in grey and 8 bits in color image.
Method 2: Key 2 [1 0 1 0 1 1 0 1] in a cover: It will embed data into the selected pixels basing on the key.
| • |
If key = [1 0 1 0 1 1 0 1] then embedding should be done in
the pixels 1, 3, 5, 6, 8. This sequence of embedding should be repeated
for a block of eight pixels for complete embedding |
| • |
If key = [1 1 0 0 0 1 1 1] then embedding should be done in the pixels
1, 2, 6, 7, 8. This sequence of embedding should be repeated for a block
of eight pixels for complete embedding |
Merits: It has a relatively lower MSE and a relatively higher PSNR.
Randomness is introduced in this method if the key is other than [1 1 1 1 1
1 1 1] as in this case it is normal raster scan LSB substitution with zero randomness.
Complexity: If the data is encrypted using DES it will introduce a complexity of (2^64).
Complexity: Good when key [1 1 1 1 1 1 1 1] is not used.
Imperceptibility: Visible to some extent if key [1 1 1 1 1 1 1 1] is used with 4-bit embedding in each pixel.
Method 3: Using Fibonacci series 1, 2, 3, 5, 8, 13 …. If the value exceeds, then use the mod length of the cover. The problem with this is collision attacks: In this method pixels are selected for embedding in the following way. A row is selected if it is not a Fibonacci number. Similarly, the column is also selected. Two methods are possible. In the first methodology, embedding can be done in all the row and column numbers that are a part of Fibonacci series. In the second methodology, embedding can be done in all the row and column numbers that are not a part of Fibonacci series.
In method 1 rows 1, 2, 3, 5, 8, … are selected and in a particular row again pixels 1, 2, 3, 5, 8, … will be selected for embedding.
In method 2 rows 1, 2, 3, 5, 8, … are selected and in a particular row again pixels 1, 2, 3, 5, 8, … will not be selected for embedding. All others pixels will be selected for embedding.
Merits: Method 1 gives more randomness and imperceptibility while method 2 gives more embedding capacity.
Complexity Analysis: If the data is encrypted using DES it will introduce
a complexity of (2^64).
For method 1: For a 256*256 image, the complexity varies with respect to size of image:
| • |
Selecting a row which is a part of Fibonacci series can be
done in (12/256) ways |
| • |
Selecting a column which is a part of Fibonacci series can be done in
(12/256) ways |
| • |
4 bits can be selected for embedding in four LSBs in 1 way |
| • |
So the final complexity of embedding 1-bit is: (2^64)*(256/12)*(256/12)*1 |
For method 2: For a 256*256 image, the Complexity varies with respect
to size of image:
| • |
Selecting a row which is not a part of Fibonacci series can
be done in (244/256) ways |
| • |
Selecting a column which is not a part of Fibonacci series can be done
in (244/256) ways |
| • |
4 bits can be selected for embedding in four LSBs in 1 way |
| • |
So the final complexity of embedding 1-bit is: (2^64)*(256/244)*(256/244)*1 |
Method 4: Variable bit embedding on the pixel (PVD). This method is
similar to embedding data into pixels based on the nature of image. If the image
is having an edge, more bits will be embedded and if it’s a smooth area
then less bits are embedded.
Merits: Since, it is taking the image texture into consideration, the imperceptibility is very high. It is also complex due to the variable bit embedding.
Complexity analysis: If the data is encrypted using DES it will introduce a complexity of (2^64).
It is taking variable bit embedding and so the complexity increases. But complexity remains the same.
Imperceptibility: Is low because it takes the image texture into consideration.
Method 5: Encrypt the secret then embed: Here, simply the data is first encrypted using some encryption algorithm and then encryption is done with the encrypted data and not with the original data. Security increases because of encryption but it will also depend on the embedding algorithm. It can be combined with any of the methods above.
Method 6: Use different encoding instead of ASCII: Here encoding is
changed for conversion of messages (may be text/audio/video) into bits. Normally
ASCII codes are used for converting a numerical data into binary. Instead of
ASCII some other encoding mechanism such as Shannon encoding, Shannon-Fano encoding,
Huffmann encoding etc can be used. Some encoding algorithms give encoded data
based on the probability of its occurrence in the data. This can increase the
embedding capacity for the same level of distortion.
Method 7: Distort the cover with known value then embed: Here the cover image will be initially scrambled and the resultant scrambled image is used for embedding of data. After embedding process is done the image is once again de-scrambled to get the stego image. By this process we can never determine the path of embedding unless the initial key for scrambling of image is known. Thus this method increases the complexity through image encryption.
Method 8: Compress the secret data then embed: Here the secret data is compressed using some reversible compression algorithm (i.e., lossless compression is performed) and the obtained data after compression is used for embedding. The complexity lies in the embedding algorithm and the compression algorithm and the latter contributes more to complexity.
Method 9: Compress, Encrypt and Embed: t includes all the above three. Compressing the data initially, encrypting the obtained result and using this final result for embedding. This gives more complexity than the previous case.
Method 10: Compress+Error control codes+Encrypt then embed: Here, the data is initially compressed by some compression algorithm and for the resultant data, error control codes such as hamming codes, cyclic codes or convolutional codes are applied for securing the compressed data. This resultant data is then encrypted and embedded. A merit of this methodology is that even if some pixels are tampered, we will be able to retrieve back our data. Complexity will be similar to the previous methodology.
Method 11: Select the pixels by Generating Pseudo random number then Embed: In this methodology embedding is done by choosing some pseudo random number. For example the pseudo random number can be generated by taking the numbers after decimal point in pi (22/7). Here embedding can be performed by taking a block of 10 pixels at a time and processing them. Consider pi as 3.141592654. Now embedding is done in 1st pixel in first block of 10 pixels and in 4th pixel in second block of 10 pixels and in 1st pixel in 3rd block of 10 pixels and so on. This type of embedding will produce randomness and increase the complexity to a very great extent.
Method 12: Select pixels based on intensity values then variable bit embedding:
In this methodology number of bits to be embedded in each pixel is decided by
the pixel intensity. A threshold is set and based on this, the number of bits
to be embedded in each pixel is decided. For example if the pixel intensity
is greater than 75 then two bits are embedded and if it is greater than 150
three bits are embedded and if it is greater than 225 four bits are embedded
or else 1 bit is embedded. Thus variable bit embedding based on the intensity
value is implemented and it increases the complexity due to its variable embedding
nature.
Method 13: Study the statistics of the cover without affecting embed: In this methodology embedding is done after taking image statistics into consideration. Here more number of bits are embedded in the edge areas and less number of bits are embedded in the plane areas and the imperceptibility is enhanced.
REVIEW ON THE STEGANALYSIS
With the aim of detecting the presence of secret messages (Qin
et al., 2010), steganalysis is classified into two categories based
on the way of looking out for the secret namely steganalysis fro specific embedding
and universal blind steganography. The specific stegananlysis is capable of
estimating the embedding ratio or even reveal the secret once the algorithm
used for embedding is known. This feature is similar to that of RS, SPA, DIH
and LSM algorithm which can detect the spatial LSB steganography and Pevni
and Fridrich (2007) algorithms which can reliably determine the presence
of secret messages embedded in the DCT domain of the image. Studies involving
the theoretical analysis and practical experiments have revealed that LSB matching
is more difficult than LSB replacing.
Though specific embedding steganography seems interesting and more fruitful, in reality, it is very difficult to implement since the method used for embedding is hardly known to the steganalysts. On the other hand, the Universal blind approaches are more attractive for practical applications since it detects the presence of secret message without the need of knowing the embedding algorithm. The flexibility and adaptability of the Universal blind approaches to get adjusted to new and previously unknown stego-methods gives it an edge over the specific methods which give comparatively more accurate and reliable results.
The Universal blind steganalysis technique (Luo et al.,
2008b) is a meta-detection method which can be adjusted to detect ant steganographic
method employed once it has been trained esing the cover and stego images. Two
types of non-specific steganalysis techniques exist in literature. The first
classifies the images as cover or stego using the original images as the training
set and to extract the features and since it does not make use o of knowledge
of the algorithms, it is purely a blind steganalysis technique. The second technique
utilizes a combination of cover and stego images as the training set. Here,
the assumption made is that special hiding techniques might have been used for
embedding but details about them is not known and hence this method is also
called as ‘half-blind’ detection method. This half-blind steganalysis
technique is an important domain of research in blind detection researches since
it has a good generalizing capability and is able to detect even new and unknown
steganographic methods.
The idea of using a trained classifier to detect secret messages has been developed
in the past years. Avcibas et al. (2003) are
the first to propose the framework for classifier-based steganalysis. Subsequently,
they presented a different feature set based on binary similarity measures for
steganalysis and its experimental results showed that, generally, SVM (Support
Vector Machines) could provide better performance than other classifiers. Farid
(2002) also has investigated the problem of classifier-based steganalysis,
and further validates that classifier-based scheme is an effective and universal
approach to deal with the twin difficulties of unknown image statistics and
steganographic algorithms. Fridrich et al. (2005)
and other researchers also studied the technique of universal steganalysis with
trained classifiers. These methods first extract an appropriate feature set
that should be sensitive to steganographic modifications and then build a classifier
to distinguish original images from stego ones.
Recently, researches show that the improved performance of the image steganalysis
is achieved at the expense of increasing the number of the feature. A few single
features for the detection of secret messages have been proposed by Liu
et al. (2008), point out that single feature is usually uncapable
of differentiating stego and cover signals effectively. And then, they extracted
two statistical properties for image steganalysis, one from the spatial domain
and the other from the DCT domain. Avcibas et al.
(2003). used ten image quality metrics as feature set and employed 18 features
from binary similarity measures of the seventh and eighth bit planes in an image
for classification Fridrich et al. (2005) extracted
23 calibrated features from DCT and spatial domain. They subsequently built
the universal classifier by exploiting 81 features which were extracted from
the higher-order absolute moments of residual noise in the wavelet domain, Farid
(2002) extracted 72 features by image decompositions based on QMF. Lately,
Fridrich et al. (2005) even have utilized 274
features for steganalysis. Chen and Shi (2008) adopted
324 features from the statistical moments of wavelet characteristic functions.
Farid and Lyu (2003). made use of a considerably larger
432 features composed of higher-order magnitude and phase statistics.
There are other works in steganalysis, For example Xia
et al. (2009) proposed a method wherein by studying neighborhood
node degree characteristics, Least Significant Bit (LSB) matching steganography
in gray images could be detected. Qin et al. (2009a)
proposed an efficient steganalytic method that detects LSB matching steganography
for the compressed and uncompressed images by constructing classifiers based
on differences between the neighboring pixels (DNPs), differences between the
local extrema (DLENs) and their neighbors in grayscale histogram. Qin
et al. (2009b) presented a survey of LSB matching steganalysis methods
for digital images. The study pondered over some persisting problems and made
way towards aspects worth researching.
However, using too many features is undesirable in terms of classification performance due to the curse of dimension. Furthermore, performing feature selection in steganalysis offers several advantages as follows: (1) Feature selection prunes the meaningless features for the classifier. (2) Feature selection may also be used to improve the classification performance. (3) Feature selection can reduce the complexity for both features generating and classifier training. (4) The selected features can help to point out the features that are sensitive to a given steganographic scheme and consequently to bring a highlight on its weaknesses. Hence, it is necessary to reduce the feature dimension by eliminating redundant features and selecting the most relevant ones.
CONCLUSION
In this review, a brief encyclopedia of infant steganography to matured steganography has been presented. Starting with the definition, differences with other security guards like cryptography and watermarking have been highlighted. Possible ways of random image steganography is the major motivation of this review and the same has been enumerated and enunciated with practical examples. As a finishing touch to the stego implementation, the significance of steganalysis has also been highlighted.
Subscribe Today
