New Text Steganography Technique Based on Part-of-Speech Tagging and Format-Preserving Encryption

Abstract

The transmission of confidential data using cover media is called steganography. The three requirements of any effective steganography system are high embedding capacity, security, and imperceptibility. The text file's structure, which makes syntax and grammar more visually obvious than in other media, contributes to its poor imperceptibility. Text steganography is regarded as the most challenging carrier to hide secret data because of its insufficient redundant data compared to other digital objects. Unicode characters, especially non-printing or invisible, are employed for hiding data by mapping a specific amount of secret data bits in each character and inserting the character into cover text spaces. These characters are known with limited spaces to embed secret data. Current studies that used Unicode characters in text steganography focused on increasing the data hiding capacity with insufficient redundant data in a text file. A sequential embedding pattern is often selected and included in all available positions in the cover text. This embedding pattern negatively affects the text steganography system's imperceptibility and security. Thus, this study attempts to solve these limitations using the Part-of-speech (POS) tagging technique combined with the randomization concept in data hiding. Combining these two techniques allows inserting the Unicode characters in randomized patterns with specific positions in the cover text to increase data hiding capacity with minimum effects on imperceptibility and security. Format-preserving encryption (FPE) is also used to encrypt a secret message without changing its size before the embedding processes. By comparing the proposed technique to already existing ones, the results demonstrate that it fulfils the cover file's capacity, imperceptibility, and security requirements.

1. Introduction

Steganography name contains two ancient Greek words: "Stegano" and "Graphy", and both relate to "Cover Writing". Steganography has been in use for decades [1]. As compared to other media like images, sounds, and video, text contains fewer redundant bits [2]. The secret data embedding in text media is the main concern in text steganography. Little modifications to the cover text will be visible as redundant bits are scarce [3]. Three characteristics are required in any steganography system: security, imperceptibility, and capacity [4][5]. A steganography system often gets to transmit hidden data via the least-covered media, which lessens the possibility of detected while transmitting through an unsafe channel. Therefore, storing secret data requires high embedding capacity [6][7]. These characteristics have the most influence on how effective a steganography setup is. However, there is generally a trade-off between the secret data capacity and the stego file quality [8][9]. For example, embedded a large quantity of secret data. In such cases, the stego files changing presents extra difficulties in terms of maintaining imperceptibility and the chance for distortion. Therefore, the primary goal of any successful proposed method is to maintain these characteristics as effectively as possible [10].

Text steganography can be divided into three classes: format-based, random and statistical generation, and linguistic [11],[13]. Firstly, in the format-based method, Text symbols' physical characteristics are employed. The changed characteristics are performed so that the human eye cannot recognize the changes [14]. In encoding the format-based method, the physical attributes of the words are changed to hide the secret data, which is dependent on symbols in different languages. Various studies focus on improving text steganography's capacity by changing the text format's physical nature using format-based approaches [15].

Secondly, in the random and statistical generation method to create cover texts, statistical features of a language are acquired. In each language, sequences of characters in a text and statistical properties, including word lengths or the frequency of occurrences of letters, are utilized in text documents [16].

Lastly, language characteristics hide secret information inside text files in the linguistic approach [17]. For example, POS can be used in English to analyze cover and secret texts and can list the definition for the text content in terms of nouns, verbs, adjectives, etc. [18]. POS is used to replace each word with a similar tag (noun to noun, verb to verb, etc.) between the secret and cover text. Alternatively, deep learning (DL) can provide an auto-generation of text used in steganography to auto-generate a cover text that already carries the secret data [19]. Consequently, the security of the cover text can be enhanced by minimizing the cover text's and stego text's visual appearance after embedding secret bits into the cover text [20]. However, Linguistic steganography raises a new challenge in security because of the semantic control of the generated texts, which are detectable by attackers.

Previous work in [9] reviewed techniques used in text steganography. One of the main issues with the embedding techniques that used non-printing or invisible Unicode characters in format-based text steganography was the position selections for embedding secret data in the cover text. Cover texts are known with limited spaces that can be used for embedding secret data, which also applies to Unicode characters. Despite these proposed approaches according to a specific language, they share a common way of selecting embedding positions inside the text: a sequential selection pattern that includes all available positions in the text document. In addition, sequentially embedding secret data in a text document where human eyes can easily detect changes negatively affects the imperceptibility and security. It makes the algorithm vulnerable to detection by a third party [21]. Therefore, many researchers used symmetric or non-symmetric cryptography techniques in secret messages before embedding them to preserve hiding capacity while providing data protection [22]. The RSA encryption algorithm was used in [23] to encrypt data to ensure security. Another study by [24] integrated steganography and cryptography for safe data transport using the Data Encryption Standard, which used a symmetric key for encryption. However, such cryptographic techniques will affect the secret data size, which increases during encryption. It negatively affects text steganography in terms of capacity and imperceptibility. Therefore, non-sequential or random embedding position techniques are preferred when embedding secret data in the cover text file. An extra layer of a security technique can be added by encrypting secret data in a way that may not increase its size, which then leads to a high-capacity text steganography system. Another issue is increasing the number of secret data bits represented by non-printing or invisible Unicode characters in text documents. As previously mentioned, these characters have limited spaces to embed secret information.

These issues motivate us to develop our proposed method based on the properties of POS tagging with non-printing Unicode Characters and Format-Preserving Encryption. In this paper, the POS tagging works as an indicator for non-printing Unicode Characters (UCs) in Format-based Steganography to improve the hiding capacity and maintain imperceptibility and security properties. The POS properties provide a complete definition for the text content by using word tags to produce a tag list. Words of the eight most common tags will be selected to establish randomized indicator positions in the cover text to hide secret data. In addition, the number of bits in each Unicode character representation is increased to four. A pair of Unicode characters will be used in the proposed method. An extra layer of security technique is applied to encrypt secret data without increasing the size by using Format-preserving encryption. So far, this paper has discussed the text steganography system's capacity, imperceptibility and security issues. One of the goals of the text steganography system is to increase embedding capacity. However, balancing imperceptibility and security is also important with the increased capacity.

The rest of the paper is organized as follows: Section 2 explains related work on text steganography, followed by POS Tagging in Section 3. Section 4 describes Format-Preserving Encryption, and a detailed description of the proposed algorithm is described in Section 5. Section 6 presents a critical analysis that includes obtained results from the experiments and a comparative discussion on former suggested methods. Finally, Section 7 discusses the conclusion and future work of the paper.

2. Related Work

This section examines related works that use the same principle concept as our proposed method, as discussed in Section 5.

A collection of forwarded e-mails is randomly selected as a cover text in a technique proposed in [15], a list of e-mails together with a randomized indexed word dictionary. The cover is an e-mail that gets forwarded. A word dictionary with a randomized index serves as the encryption key for the concealed data in the carbon copy field. A random bitstream (temporary stego key) using the system's time is generated and sent independently using public-key cryptography. The index values of the dictionary terms are randomly generated using this temporary stego key. This approach is safe and secure against typical attacks since it removes noise from the actual e-mail body information. Additionally, it provides an extra degree of protection by randomly generating keys using the system time for the word index values.

The linguistic text steganography method is based on POS tagging proposed by [18]. This method processes secret and text covers with POS tagging to generate a tag list of each secret and cover text. The embedding process replaces each word from the cover text with a word from the secret text based on similarity tags, such as (noun with noun), (verb with verb), etc. Although the cover text remains unchanged in terms of the formatting and the level of similarity is high, it is difficult to control the semantics of the cover text after the embedding process (which is to keep the text's original meaning). At the same time, capacity is limited since the replacement condition is finding the same tag from the secret and cover text.

A multi-keyword coverless text steganography algorithm based on pre-treatment and POS tagging was proposed by [19]. The author used the Chinese Mathematical Expression that represents the Chinese characters. Chinese characters are divided into about 644 basic components. For example, the Chinese character "丛" consists of two parts, "人" and "一". The embedding process starts by dividing the secret message into sequence keyword segments and then randomly changing the sequence. The next step marks the locating keywords using Chinese character Mathematical Expression components. Then, POS is used for embedding the number of keywords inside the cover text and mapping their locations as a key. On the receiver's side, stego-texts are received, and secret messages can be extracted using the mapping locations and random sequence keys.

The author in [25] suggested a text steganography method based on a set of two-letter words from the Oxford Dictionary. This method used a set of two-letter words as markers to secret data embedding. Every two secret text bits are mapped to a certain non-printing Unicode character (UC). Through this procedure, a shared UC mapping table between the sender and recipient is created. The work in [26] suggested a text steganography method based on the Lempel-Ziv-Welch Algorithm and sets of two-letter words that improve the previous method in [25]. The Lempel-Ziv-Welch (LZW) compression algorithm compresses the secret text to minimize its size. A colour-coding approach is used with the LZW compression technique in [27]. The technique used forward mail as a cover medium to obfuscate secret information. The operation first compresses the secret information before disguising it in the e-mail addresses and cover messages. The cover text (or message), which is coloured using a colour-coding table, has the secret data bits included in it. The study's findings showed that the strategy had a higher embedding capability than previous approaches and was less computationally difficult. Additionally, using stego keys greatly improves the security of the suggested method.

A text steganography method in [28] uses e-mail addresses with Huffman compression to improve the capacity of the cover file. The length of an e-mail ID character is used to hide the bits of hidden messages.

Table 1 summarizes the related works of text steganography and their strong and weak points.

Table 1. Related work on text steganography

3. Part-of-Speech Tagging

Before discussing the proposed model, this section explains the related concepts of POS Tagging. POS tagging is applied to create a list that contains tags of the most common POS categories in English in the cover text file. POS tagging uses morphological analysis to add a POS to each word in a sentence. Morphology is the science of word-forming, or how words are constructed from smaller components. Tokenization, dictionary lookup, and disambiguation are the processes that make up morphological science [18]. A dictionary lookup takes a string and returns a list of lexemes with POS data. In general, there are two types of POS tagging: supervised and unsupervised tagging. A pre-tagged corpus is needed for constructing a POS tagger tool in supervised tagging, while in unsupervised tagging, no pre-tagged corpus is required; instead, advanced computational methods are used to produce a tag set automatically. The POS tagging process starts by breaking down the given text in the natural language processing into the smallest unit in a sentence called a token. Punctuation marks, words, and numbers can be considered tokens [19]. The next step is POS tagging, which labels each word in a sentence with its appropriate POS.

The Penn POS tag set containing 36 Penn sets is shown in Table 2 [18]. Common POS categories in English are eight Penn tag sets (noun, verb, adjective, adverb, pronoun, preposition, conjunction, and interjection) [29]. These eight POS Penn categories in English are chosen as an indicator among 36 Penn sets of POS for hiding data in the proposed work, which is discussed in the next section.

Table 2. Penn POS Tag Sets

Following is an example with an input text for a POS tagger and the output after using POS tags from Table 2.

Input:

“The transmission of confidential data using cover media called steganography.”

Output:

Step 1: text tokenization: “The, transmission, of, confidential, data, using, cover, media, called, steganography”

Step 2: text tagging: “The DT, transmission NN, of IN, confidential JJ, data NNS, using VVG, cover NN, media NNS , called VVN, steganography NN”

4. Format-Preserving Encryption (FPE)

The size-increasing problem after encryption processes can be addressed with a unique symmetric encryption technique known as FPE, which is growing acceptance. Compared to well-known encryption methods like AES and DES, this approach is slightly different. [30]. It is a rapidly developing tool for cryptography that ensures data security. Format-preserving encryption seeks to encrypt data without modifying its size or format. Therefore, data will be encrypted differently than traditional encryption, ensuring that the output is the same size and format as the input. There are only two operating modes that the National Institute of Standards and Technology (NIST) recommends: FF1 and FF3. It is suggested to employ the Cipher Block Chaining (CBC) operation that has the ability to encrypt messages of any length using a fundamental block cypher component known as BPS-BC. FF1 and FF3, which stand for BPS-BC (Brier Peyrin Stern) and format-preserving Feistel-based encryption, respectively, were suggested by [30] and [31]. Any format of data blocks can be encrypted using the electronic code book (ECB) operating mode. Even though AES is utilized as the underlying block cypher, operating modes may be considered as a type of FPE block cypher. FPE can also be employed in communication systems when it is necessary to encrypt certain protocols, such as in industrial or military situations or when encrypting specific media [32]. In this paper, in order to improve the security level of stego text files while maintaining the amount of capacity and imperceptibility, a new approach is proposed that focuses on employing format-preserving encryption FF1 mode with POS tagging in a format-based method.

5. Proposed Method

In this section, our proposed method is explained in detail, which consists of three phases: (1) Building a POS Tagging List, (2) Embedding, and (3) Extracting secret messages.

5.1 Building a POS Tagging List

This phase produces outputs: a POS Tagging List (for cover text). The output of this phase is converted into inputs for the embedding phase after checking the embedding capacity of the cover text. Eight Penn tag sets will be used as indicators: nouns, verbs, adjectives, adverbs, pronouns, prepositions, conjunctions, and interjections [29]. Fig. 1 shows Algorithm 1, which is to establish the POS list.

Fig. 1. Algorithm to create the list of eight most common POS tag tokens.

Algorithm-1: Building a POS tagging list

Input: Cover text file (C)

Output: List of eight most common POS tag tokens

5.2 Embedding Phase

This phase proposes a text steganography technique that exploits the English text redundant data, using POS to resolve the cover text file's limited space for hiding data. This method uses the eight most common POS tags as indicators for hiding secret text data in English scripts using non-printing UCs. Using POS provides randomization in selecting indicators in specific positions in the cover text for UCs representing data bits. Then, FPE is applied to the secret data. FPE is a technique used to encode plaintext in such a way that it can preserve its original length and format. Using FPE can increase the capacity to embed data while simultaneously maintaining the steganography system's security and imperceptibility.

The algorithm in Fig. 2 (Algorithm 2) uses UCs to hide eight bits of a secret message after each indicator word in the cover text. In Algorithm 2, the cover text is read at the first step and checked for POS tags. Next, create a list of tokens (words) and their tags of the eight most common POS in the text cover. Then, convert the secret message to binary, encrypt the secret message using FPE and divide the binary into blocks of 4 bits. After that, check every 4 bits according to the non-printing UCs mapping defined in Table 3. Finally, insert a pair of UCs (two non-printing Unicode characters) after each word (token) of a specific tag according to the eight most common POS tag lists to hide 8 bits in each position. The operation starts with tokens set as noun tags and then moves to the next tokens with another tag until all eight tags tokens in the list are finished. Therefore, the randomization concept will be achieved in specific positions in the cover text since the tokens for each tag are already located and randomly distributed based on the sentence of the cover text. Fig. 3 shows the Embedding phase flowchart using the proposed method.

Table 3. Non-printing UCs map for hiding 4 bits in each character

Fig. 2. Algorithm for Embedding data using the proposed method.

Fig. 3. Flowchart of embedding data using the proposed method

5.3 Extraction phase

POS is also applied to retrieve the secret message encoded from the stego text file. Reading the stego text initiates the extraction process. to check for POS tags and create a list of tokens (words) and tags of the eight most common POS in the text cover. The next step is extracting the alternative Unicode character after each POS tag and mapping each UC according to the UC map defined in Table 2 to show 4 bits of hidden data. After that, Fk of FPE is used to decrypt the bits of the secret message. Finally, the bits that are recovered are decoded and converted into a letter to reconstruct the secret message text (original). Fig. 4 shows Algorithm 3, which is the extraction process of the proposed method.

Fig. 4. Extraction process of the proposed method.

6. Experimental Analysis and Results

This section evaluates the experimental results of our proposed method. Capacity, imperceptibility, and security are used to evaluate how well the proposed method performs as follows:

6.1 Imperceptibility analysis

In steganography, similarity helps in achieving the imperceptibility criterion. Imperceptibility is one of the most important requirements of steganography systems. It refers to the system's ability to avoid detection. The features of the human visual system (HVS) underpin this necessity. Similarity metrics and the Jaro–Winkler distance can measure a quality stego text's transparency. Similarity metrics analyze the number of identical characters and the transposition of characters in two strings.Jaro-Winkler metric analyses the number of identical characters and the transposition of characters in two strings [18]. The original cover text is compared before and after the embedding procedure to see if the proposed POS method can produce high-quality stego text or if the text's visibility is affected. Equations that calculate the Jaro–Winkler value, which is used to assess the quality of the stego text, are as follows:

Jaro_Winkler(S, C) = Jaro_Score + (L * P * (1 - Jaro_Score))       (1)

\(\begin{align}Jaro\_Score=\frac{1}{3}\left(\frac{m}{\operatorname{Length}\left(s_{1}\right)}+\frac{m}{\operatorname{Length}\left(s_{2}\right)}+\frac{(m-t)}{m}\right)\end{align}\)       (2)

Where L is the length of the common prefix at the start of the string up to a maximum of 4, P is the constant scaling factor (0.1 ≥ P ≤ 0.25), s₁ is the initial string, s₂ is the second string, m is the number of matched characters, and t is the number of transpositions.

As indicated in Table 4, the proposed POS method is tested on secret messages embedded in a 10,768-bit cover text file. This setting was used in [25][26].

Table 4. Jaro–Winkler experimental results for the proposed methods

The Jaro–Winkler values of 11 secret messages hidden in the cover file and the stego cover files generated by the proposed method are shown in Table 4. The Jaro–Winkler value of 98.8% is achieved by the stego cover text files. However, because each Unicode non-printing character takes three bytes to represent in the text, the size of the stego file will be increased. The performance of the proposed POS method is compared based on the quality of the stego cover to that of techniques used in previous studies [25] and [26]. Both papers used a format-based technique with a sequential embedding pattern by embedding secret bits after each two-letter word, using all available positions in the cover text with UCs. The proposed method gives 95.3% compared to the TWL method in [25] and 98.3% with the LZW method in [26]. The comparison focuses on the stego cover's average Jaro–Winkler value. Fig. 5 presents comparison results, where the proposed method outperformed the other two.

Fig. 5. Comparison of the Jaro–Winkler values of the proposed with [25] and [26] methods

The proposed method was also compared to the results from [18] in terms of the quality of the stego cover. The technique in [18] also employed POS Tagging in linguistic text steganography. Fig. 6 presents the comparison result, with the proposed method performing better than the linguistic (POS) method.

Fig. 6. Comparison of the Jaro–Winkler values obtained by the proposed and [18] method

The comparisons aim to show that the proposed POS method with UCs in the format-based method can achieve a higher result than [25] and [26], which are also based on format-based steganography. Meanwhile, the comparison in [18] was performed in linguistic steganography. Techniques in [25] and [26] insert a higher number of UCs in all available positions in a sequential pattern. In contrast, the linguistic (POS) technique in [18] replaces each word from the secret message with another word from the cover text file, minimizing the similarity between cover and stego text, which is used to count Jaro–Winkler values.

6.2 Capacity analysis

The proposed method is evaluated in terms of capacity ratio and hiding capacity ratio as follows:

6.2.1 Capacity ratio (CR)

Capacity refers to the ability of the cover text to embed secret data. CR can be measured based on the hidden data (bytes) embedded in the cover file. This metric can be mathematically expressed as follows [33], [34]:

\(\begin{align}\text {Capacity Ratio}(C R)=\frac{\text { Size of the embedded data }}{\text { Size of the coverfile }} \times 100\end{align}\)       (3)

An evaluation is performed in terms of its CR using a set of English script resources obtained from http://www.textfiles.com for the first 18 experiments and four experiments from previous studies. In Table 5, experiments 19 and 20 are from [26], and Experiments 21 and 22 are from [18].

Table 5. Experimental CR results for the proposed POS method

Capacity significantly depends on the embedding technique. Since every four bits of the secret message are represented in each UC, the POS method embeds eight bits into each POS tag word through a pair of UCs. Table 5 presents the CR of the proposed POS method.

The proposed method obtains a higher CR than the method proposed in the literature [26], which produces a slight variation in the size of invisible characters and hides a maximum of 117 and 96 bytes in the cover files. In comparison, the proposed method can hide (203 bytes and 186) in the same cover files (Experiments #19 and #20 in Table 5). Fig.7 presents the comparison between the two methods.

Fig. 7. Comparison in CR between the proposed method and [26]

As explained before, experiments #21 and #22 use POS Tagging in linguistics to hide text [18]. Replacing each word from the secret message with another word that must have the same tag as the cover text will limit the capacity to hide the secret message since the replacement condition uses the same tag from both the secret and cover text. The embeds of [18] are 21 and 27 bytes. Meanwhile, the proposed method embeds 128 and 78 bytes into all space characters, as shown in Table 5. Fig. 8 compares the two cover files used in experiments #21 and #22 and the number of hidden data bits.

Fig. 8. Comparison in CR between proposed method, and [18]

6.2.2 Hiding capacity ratio (HCR)

A key determining factor in evaluating a text steganography method is hiding capability. The hidden message's ratio size to the stego text's size is known as the hiding capacity HCR. HCR can be mathematically expressed as follows [26],[28]:

\(\begin{align}\text {Hiding Capacity}(H C R)=\frac{\text { Size of the seret message }}{\text { Size of the stego file }} \times 100\end{align}\)       (4)

The proposed method is evaluated in terms of its HCR, as shown in Table 6. In the experiment, secret messages of different sizes were embedded in a cover text file.

Table 6. HCR of the proposed method using 13 secret messages

From Table 6, the proposed method has an average HCR value of 3.35%. Increasing the size of the secret message can also increase the HCR value. Fig. 9 compares the secret message size and HCR using the proposed method.

Fig. 9. Comparison between the secret message size and HCR using the proposed method

The proposed method is also compared with several recently proposed approaches, as discussed in the Related Work section. The same resources used in previous studies are applied for the comparison, as shown in Fig. 10 and Fig. 11. The HCR value of the proposed method obtains a capacity of 14.21%, which is superior to the methods proposed by previous studies. Fig. 12 presents the HCR comparison of the proposed method with previous studies.

Fig. 10. Secret message

Fig. 11. Cover message

Fig. 12. Comparison (HCR) of the proposed method compared to previous methods

6.3 Security Analysis

Security is one of the important steganography requirements. The security ratio will be small in any text steganography method if excess characters are raised more than the original characters in the cover text [20]. The security ratio is calculated based on the number of increased letters, visible or invisible in the stego-text, compared with the original text. The following equations from [34] are used in the security evaluation. Table 7 presents the security ratio obtained by the proposed method.

Table 7. Security evaluation of the proposed POS method

\(\begin{align}\text {Discount value}=\frac{\text {Excess character }}{\text { Original character }} \times 100\end{align}\)       (5)

Amount after discount = Original character − Discount value       (6)

\(\begin{align}\text {Security ratio}=\frac{\text { Amount after discount }}{\text { Original character }} \times 100\end{align}\)       (7)

Table 7 shows the security ratio values of 11 secret messages hidden in the cover file and the generated stego cover files through the proposed method. The stego cover text files achieve a high-security ratio value of 99.29. The average security ratio of the stego file is high since the Unicode characters used in the embedding process represent 4 bits of the secret message. The use of randomized position in the proposed method is based on eight tags only, which means that not all available positions in the cover text will be filled with Unicode characters. Therefore, the number of UCs used will be less even when used in pairs in the proposed method.

7. Conclusion and Future Work

This paper proposes a new format-based text steganography technique using POS tagging and format-preserving encryption. The aim is to use the randomization concept in selecting positions with encryption for data hiding without increasing the secret message size. Therefore, the Unicode characters can be used more efficiently for embedding in cover text. The proposed method uses a 4-bit secret message represented in each Unicode character. A pair of Unicode characters are inserted after each specific tag token in the cover text, providing high embedding capacity and efficiency with minimum effects on imperceptibility and security. Results show that the proposed method has demonstrated improvement in capacity, security, and imperceptibility compared with previous studies. For future research, enhancement can be made by decreasing the secret message size using compression techniques to further increase the capacity of the steganography system.