INTRODUCTION

Steganography aims to hide the very presence of covert communications by imperceptibly embedding secret information into innocent-looking carriers, such as digital images, videos, texts, etc. With the increasing use of computer and the development of Internet, text is becoming increasingly popular in daily life. This motivates the development of linguistic steganography, which refers to employing the knowledge of natural language processing for camouflaging information into texts (Atallah et al., 2000).

Linguistic steganography has formed a new research field in steganography, which has recently attracted the attention of many researchers (Bergmair, 2007). The contemporary linguistic steganography in the literatures can be grouped into three categories in terms of the way of generating the stego texts.

The first class is generalized to the approach that directly produces a new natural-like stego text without the support of a given cover text. The primary methods mainly based on mimicking technique to convert the secret information into a text that obeys the statistical properties of another particular normal natural language text. In order to improve the readability and quality of the produced texts, probabilistic context-free grammars are combined with mimicking technique to generate stego texts (Wayner, 2002). Despite the fact that the generated texts have almost regular grammar and syntax, their contents are always meaningless and implausible to human readers. The transmission of these texts between different communication parties may raise suspicion by people inspection. Moreover, the existence of communicated secret information can be detected by using statistical characteristics of correlations between the general service words (Chen et al., 2008).

The second class which converts the cover text into a complete different normal-like stego text, can be considered as an extension to the first class mentioned earlier. Combining text-mimicking technique with PPT documents, Liu et al. (2008) converted secret information into innocuous sentences by exploiting the resource from a cover PPT document. Then the generated sentences are written into the note pages of the cover document to enhance the security. The efficiency and security of this steganography is greatly improved compared with the methods of the first class. Another typical method of this class is based on machine translation (Grothoff et al., 2005; Stutsman et al., 2006; Grothoff et al., 2009). Its critical point is to create multiple translations for each sentence in cover text and then select one of these translated sentences to encode information. The generated stego texts with the help of basic information provided by cover texts have better readability and quality than those produced by steganography of the first class. However, they also have a disadvantage of causing suspicious from text steganalysis (Meng et al., 2010). Steganalysis is the art of discovering secret information in digital media. Once a steganographic method is defeated by steganalysis, the secret information is no longer secure.

The approaches of slightly modifying an existing text to obtain a stego text are defined as the third class. This class of methods manipulates lexical, syntactic or semantic properties of a given text while preserving the meaning as much as possible to embed information (Topkara et al., 2005). Previous research on linguistic steganography is mainly focused on the approaches employing synonym substitution and syntactic transformation for hiding information.

The synonym substitutions-based approach embeds information in the way of replacing the word by its synonym to represent designated value. In theory, the modifications preserve the meaning of cover texts. However, simple synonym substitutions may cause potential wrong syntax and semantics since the synonyms with the identical or similar meanings still have semantic and pragmatic differences. Some more superior methods were proposed to overcome the vulnerability of synonym substitutions (Bolshakov, 2004; Topkara et al., 2006a). Bolshakov (2004) proposed to directly substitute an absolute synonym for a word and previously test a relative synonym whether it is semantic compatible with collocations containing the replaced word. Topkara et al. (2006a) used a quantitative resilience criterion to just choose one most appropriate alternative for every being replaced word in a text. However, the above efforts were just focused on improving the compatibility of the synonym substitutions with the context and do not consider the changes of the inherent statistical characteristics of synonyms sequence, such as occurrence frequency of the synonyms. The synonym substitution-based steganography has low resistance of the existing text steganalysis attack (Taskiran et al., 2006; Yu et al., 2008).

Another extensively concerned steganographic approach of the third category is based on syntactic transformations (Atallah et al., 2001; Liu et al., 2005; Wang et al., 2008; Topkara et al., 2006b; Murphy and Vogel, 2007; Meral et al., 2007; Meral et al., 2009; Kim, 2008). This approach expresses a sentence into a different semantically equivalent sentence structure for hiding information. For instance, it transforms an active sentence into a passive one without any serious meaning and grammaticality change. Atallah et al. (2001) were the first ones to employ transformations described in post-Chomskian generative syntax to design a natural language watermarking scheme. Topkara et al. (2006b) attempted to resolve the implementation challenges syntactic transformations-based steganography and designed an objective evaluation metric to evaluate the quality of the stego text. By taking the characteristics of the language into consideration, several sophisticated steganographic methods have been proposed using relevant particular syntactic transformations of different languages such as Chinese (Liu et al., 2005; Wang et al., 2008), Turkish (Meral et al., 2007; Meral et al., 2009), Korean (Kim, 2008), etc. Comparing with synonym substitution-based method, the syntactic transformation-based method has high robustness to resist attacks but it also demands deep structure analysis and its embedding capacity is limited. In practical, a perfect steganography tool is hard to be implemented and widely used for lacking enough natural language knowledge and matured reliable tools of natural language processing.

Linguistic steganography is gradually improved recently but much more efforts are need to avoid linguistic flaws in stego texts. However, it is a difficult task of building a complex natural language processing systems. Some linguistic steganographic methods possess of low popularization and application value or the stego texts in practical always have degradation of quality in theory. And texts always provide few redundancies of linguistics for embedding information leading to low embedding bit rate of linguistic steganography. Moreover, most steganographic methods have been defeated by using the distortions of statistics with the development of text steganalysis (Taskiran et al., 2006; Chen et al., 2008; Meng et al., 2010; Yu et al., 2008). These vulnerabilities and concerns of the above linguistic steganography motivate to develop a linguistic lossless steganographic method with high resistance of steganalysis attacks and improved embedding bit rate as well.

Multiple choices are the most popular and frequently used form in educational testing, professional certification examination and many other sorts of tests and surveys. Employing MCQs to hide information would not arouse suspicious from eavesdroppers. Considering the characteristics of MCQs, a secure steganography is proposed by taking MCQs as the cover carriers in this study. Firstly, by further investigation on MCQs, two inherent attributes are introduced for steganography, because of their peculiar nature. Then each MCQ is assigned to two determinate attribute values based on the character lengths and the order of its options. Selecting MCQs from a prepared MCQs bank according to their first attribute values will generate a stego text. And modifying the second attribute values of MCQs via equivalent transformations would embed more information.

MCQS-BASED STEGANOGRAPHIC METHOD

Within today’s society, various examinations or tests designed to measure people’s knowledge, skills, aptitude, etc., in many different fields such as the education, professional certification, psychology, the military, etc. MCQs are the basis of a significant portion of a test. In a test that has items formatted as MCQs, a candidate is given a number of set answers for each question and the candidate must choose which answer or group of answers is correct. With the convenience brought by Internet, many websites offers abundance MCQs for personal research and private study. Especially, some special MCQs banks of many subjects are provided. Before attending a formal examination, some exercises are always done by using these materials. Communication via a text including MCQs retrieved from a question bank is commonplace, thus it is convenient to transmit secret information by using the MCQs as the cover mediums and could not arouse suspicious from eavesdroppers.

Attributes of MCQs: A MCQ always consists of a stem and a series of options. A test requires a test taker to choose all appropriate answers or only one answer from a list of options. Two vocabulary MCQs of primary English are provided in the following as examples:

MCQ examples:

It is worthwhile to note that the MCQs in a test always serve the same subject. In many formal standardized tests, a test question is often retrievable from a question bank. From the appearance point of view, the arrangement of MCQs in a test paper is always automatically made; the order of MCQs can be adjusted based on requirements. On the other hand, options of well-written MCQs are independent of one another and consistent in logic and grammar to the stem. Therefore, the order of options can also be arbitrarily changed. With these characteristics, MCQs can be successfully used to hide information. Before describing the proposed steganography, two attributes, namely Q-attribute and P-attribute are introduced respectively for each MCQ.

Given a MCQ t = <a₁,a₂,...,a_n>, where a_j denotes the jth option. Let L(a_j) denote the character length of the option a_j. After rearranging the options in an increasing lexicographic order, the options are denoted as a new vector <a'₁,a'₂,...,a'_n> . According to the character lengths of the reordered options, a Q-attribute denoted as g(t) is defined and calculated by:

(1)

where, b_j = L(a'_j) mod 2.

Apparently, g(t) is a random n-bit integer, whose value is in range of 0 to 2ⁿ. For the example MCQ 1 t₁ = <called in, called for, called off, called out>, after rearranging the options in an increasing lexicographic order, a new vector of options <called for, called in, called off, called out>, is obtained, then, L(a'₁) = 10 L(a'₂) = 9, L(a'₃) = 10, L(a'₄) = 10, thus its Q-attribute value is 4. In the same way, Q-attribute value of the example MCQ 2 figured out at 14. MCQs’ Q-attribute values would be approximately independent and identically distributed (i.i.d.) in a large MCQs bank, since the character length of options in a MCQ is random and all options are mutually independent. MCQs with appointed Q-attribute values can be successfully retrievable from a large MCQs bank for embedding information.

Based on above analysis, we can also find that the options of a MCQ can be arranged into a required layout. Each possible arrangement can be assign to a unique value. For a given MCQ, the arrangement of its options will map to a determined value. Given the MCQ t = <a₁,a₂,...,a_n>, its n different options have n! number of permutations. Arranging the n! permutations in a certain order (lexicographic order, for example), the position of <a₁,a₂,...,a_n> locating at is defined as the MCQ’s P-attribute, denoted as f(t). In this study, we define the position starts at 0. The P-attribute value is a integer in range of 0 to n!-1. Generally, the options in a MCQ are always randomly placed in practical tests. Thus, the options can be reordered to make the MCQ’s P-attribute equal to a required value in a certain range without influencing the meaning of the MCQ and bringing any errors.

Next we illustrate the procedure of obtaining the P-attribute value of a MCQ. Set the n options of MCQ t = <a₁,a₂,...,a_n> in increasing lexicographic order is a'₁,a'₂,...,a'_n. For convenience, we set n = 4 forexample.

Fig. 1:	Diagram of embedding process

The 24 permutations in increasing lexicographic order are listed as follows:

Each permutation has its own position in the ordered permutation set. From <a'₁,a'₂, a'₃,a'₄> to <a'₄,a'₃,a'₂a'₁>, they are numbered 0 to n!-1, respectively. For a MCQ t = <a₁,a₂,a₃,a₄>, if it is happen to a₁ = a'₁, a₂ = a'₂, a₃ = a'₃, a₄ = a'₄, then its P-attribute f(t) = 0. If a₁ = a'₁, a₂ = a'₃, a₃ = a'₂, a₄ = a'₄, then f(t) = 2. For example MCQ 1 t₁ = <a₁,a₂,a₃,a₄>=<called in, calledfor, called off, called out>, after rearranging the options in an increasing lexicographic order is obtained <a'₁,a'₂,a'₃,a'₄>=<called for, called in, called off, called out>, then a₁ = a'₂, a₂ = a'₁, a₃ = a'₃, a₄ = a'₄, thus f(t₁) = 6. In the same way, the P-attribute of example MCQ 2 is calculated to 9.

The embedding process: With the help of the two attributes of MCQs, secret information can be securely embedded into MCQs. Before embedding secret information, we collected enormous MCQs to form different banks based on their subjects. In the embedding process, according to the secret information and the options number of a MCQ in the collected MCQs bank, two integer sequences will be calculated and encoded into MCQs. MCQs whose Q-attribute values orderly equal to the first integer sequence are selected to generate a stego text. Then, the options of the selected MCQs are reordered to make their P-attribute values accord with the second integer sequence. Finally, the generated stego text consisting of selected and modified MCQs is output. Diagram of embedding process is illustrated in Fig. 1.

Now, we describe the embedding process of our MCQs-based steganography in detail:

Extraction process: In the extraction process, a larger integer is first obtained from the two attribute values of

MCQs in the stego text. Then translating the integer into decrypted bytes can retrieve the secret information. The Extraction process involves the following three steps:

(2)