Steganographic text steganalysis resistance enhancement method based on deep learning

Abstract

The application belongs to the technical field of information hiding in information transmission, and discloses a steganographic text anti-steganalysis capability enhancement method based on deep learning. The method learns the characteristics of the data set in the open domain environment by using the deep learning method, generates steganographic text with the statistical distribution of the original data set while embedding secret information, and finally recommends the corresponding emoticons for the steganographic text according to the distribution of the emotional tendency and emoticons of the steganographic text in the real environment, so as to further strengthen the anti-steganalysis capability of the steganographic text. The application solves the problem of missing emotional clues in the generated text steganography, so that the anti-attack capability of the steganographic text is stronger when the steganographic text is transmitted in the public channel.

Description

Technical Field

[0001] This invention belongs to the field of information hiding technology in information transmission, and specifically relates to a method for enhancing the anti-steganography analysis capability of steganographic text based on deep learning. Background Technology

[0002] The field of information hiding mainly studies how to efficiently and securely embed secret information into other information carriers, with the aim of ensuring its security by concealing its existence.

[0003] In recent years, with the rapid development of the internet, more and more people are using it for information transmission and exchange. This has shifted the main battleground for protecting confidential information from specific scenarios to open domain environments. In practical terms, generative text steganography, compared to retrieval and modification text steganography, has advantages such as the ability to generate steganographic text that conforms to the statistical distribution characteristics of the real-world carrier, maintaining a certain level of concealment while offering a larger embedding capacity, high practicality, and low noise. Therefore, it is more suitable for open network environments. Social media is filled with emoticons used to reinforce the emotions and clarify the meaning of sentences. Emoticons serve as emotional and visual cues to help users understand their intentions, greatly reducing ambiguity in the recipient's understanding of the sender's intent. This has led to their increasing popularity on social media, with emoticons accounting for approximately 20% of online text on common domestic and international social platforms. However, similar studies have not considered the human visual perception issues caused by emoticons, making this issue crucial. In previous studies, researchers have failed to fully consider the distribution characteristics of real channels. They have ignored the existence of emojis in real channels (i.e., emojis were removed from the text during the preprocessing stage). This inevitably leads to an excessively large statistical distribution between the steganalysis carrier and the real carrier, resulting in a decrease in the steganalysis resistance of the stegtext, that is, a decrease in the ability to protect secret information, and making the covert system more insecure. Summary of the Invention

[0004] To address the issue that previous studies have failed to fully consider the statistical distribution characteristics of real channels, this invention proposes a deep learning-based method to enhance the anti-steganography capabilities of steganographic text. This method fully considers the characteristic statistical distribution in real channels and further improves the anti-steganography capabilities of steganographic text in public channels.

[0005] The objective of this invention is achieved through the following technical solution: a method for enhancing the anti-steganography capabilities of steganographic text based on deep learning, the method comprising the following steps:

[0006] A deep learning-based method for enhancing the steganalysis resistance of steganographic text includes the following steps:

[0007] Step 1: Collect online text from social media platforms;

[0008] Step 2: Perform appropriate preprocessing on the dataset, including removing special characters, hyperlinks, and phone numbers to form the dataset;

[0009] Step 3: Collect statistical distribution features of emojis in the data set and classify them according to semantics, and design two emoji recommendation algorithms: Insert algorithm and Replace algorithm;

[0010] Step 4: Further process the dataset for training the stegtext generation model and the emoji prediction model, respectively;

[0011] Step 5: Use the stegtext generation model trained in Step 4 to generate stegtext with secret information, and input the stegtext into the emoji prediction model trained in Step 4. Add emojis to the stegtext according to the recommendation algorithm in Step 3.

[0012] Step 6: Calculate the basic indicators of the three sets of data: the plain text steganography generated by the steganography generation model in Step 5, the steganography with emojis added by the Insert algorithm, and the steganography with emojis added by the Replace algorithm. Input these indicators into the steganography analysis model to compare the resistance to steganography analysis of the three sets of steganography.

[0013] Further, step 1 includes the following steps:

[0014] Step 1-1: First, use the general stop word list of the corresponding social media platform as the keyword library for the web crawler;

[0015] Step 1-2: Randomly select a stop word from the stop word library as the keyword for one round of data crawling;

[0016] Steps 1-3: Customize the number of keywords crawled in each round, aiming to statistically analyze the characteristics of the communication channel itself.

[0017] The data volume is no less than 1 million records.

[0018] Furthermore, step 3 includes the following steps:

[0019] Step 3-1: Statistical analysis of emoji frequency characteristics: Statistical analysis of the frequency of each emoji and the percentage of online text containing emojis;

[0020] Step 3-2: Calculate the percentage of online text containing punctuation marks;

[0021] Step 3-3: Analyze the location information of emoticons in online text, dividing them into sentence beginning, sentence middle, and sentence end;

[0022] Steps 3-4: Calculate the percentage of online text containing each number of emojis;

[0023] Steps 3-5: Based on the statistical information, design an emoji recommendation algorithm for macro-control.

[0024] Furthermore, steps 3-5 include the following steps:

[0025] Step 3-5-1: Algorithm Parameter Description:

[0026] A: The set of plain text steganography generated by the steganography generation model in step 5;

[0027] s: A single line of steganographic text in A;

[0028] s1: Steganographic text after the emoji is inserted;

[0029] B: The steganographic text set after adding emojis in step 5 using the emoji recommendation algorithm;

[0030] f: The percentage of online text containing emojis in the dataset obtained in step 3-1;

[0031] f1: The percentage of online text containing punctuation marks as calculated in step 3-2;

[0032] P: Location information of emoticons in online text obtained in step 3-3;

[0033] p: The specific value in P, which can be one of the following: the beginning of the sentence, the middle of the sentence, or the end of the sentence;

[0034] P1: A list of punctuation mark positions in the current web text;

[0035] p1: A specific value randomly selected from P1;

[0036] N: The percentage of online text containing each number of emojis as calculated in steps 3-4;

[0037] n: A specific value in N, with a range of n∈[1,5], where n is a positive integer;

[0038] Step 3-5-2: Design an emoji insertion algorithm:

[0039] 1) Select s from A, and then determine whether to insert an emoji based on the probability value of f. If the result is False, proceed to 2); if it is True, proceed to 3).

[0040] 2) Output s to B, and the steganography operation ends;

[0041] 3) Select the quantity information n from N as the number of emojis to be inserted into s;

[0042] 4) The emoji prediction model performs sentiment prediction on the input sentence, then selects an emoji from the predicted emoji classification pool based on the true statistical distribution, and then selects the position information p from P as the insertion position of the emoji in s. This process is repeated n times.

[0043] 5) After completing operation 4), obtain s1 and output s1 to B;

[0044] Step 3-5-3: Design an emoji replacement algorithm:

[0045] 1) Select s from A and determine whether s contains punctuation marks. If the result is True, proceed to 3); otherwise, proceed to 2).

[0046] 2) Output s to B, and the steganography operation ends;

[0047] 3) According to The probability value is used to determine whether to perform an emoji replacement operation. If the result is True, proceed to step 5); if the result is False, proceed to step 4.

[0048] 4) Output s to B, and the steganography operation ends;

[0049] 5) Select the quantity information n from N as the number of emojis to be inserted into s;

[0050] 6) Calculate the punctuation mark position index P1 of the current steganographic text, and randomly select a replacement position p1 from it;

[0051] 7) Use the emoji prediction model to make n predictions for s, and merge the n emojis into a string to replace the punctuation mark at position p1;

[0052] 8) After completing operation 7), obtain s1 and output s1 to B.

[0053] Furthermore, step 4 includes the following steps:

[0054] Step 4-1: Remove all emojis from the dataset to obtain a dataset in pure English text format for training the stegtext generation model;

[0055] Step 4-2: Extract all network text containing emojis from the dataset for training the emoji prediction model. Further, step 4-2 includes the following steps:

[0056] Step 4-2-1: Based on the statistical results in Step 3-1, construct an emoji library S. The construction criterion is to select emojis that account for more than 90% of the total number of emojis.

[0057] Step 4-2-2: Divide the emojis in the emoji library into m categories according to their semantics, and assign a number and Huffman coding to each category, i.e.:

[0058] S = {c1, c2, ..., c m}

[0059] c i ={h1,h2,…,h n}

[0060] Where m represents the m categories of emojis in S according to their semantics, and c m c represents the m-th class; i Let h represent the i-th class (2≤i≤m) in S, where h n This represents the Huffman code of the nth emoji in this class;

[0061] Step 4-2-3: Construct data labels.

[0062] Furthermore, step 4-2-3 includes the following steps:

[0063] 1) Calculate the frequency of emoticons in a data point, sort them in descending order to form a list;

[0064] 2) When the length of the list is equal to 1, determine if list1 belongs to c in S. i If the data is of class i, then the data entry is labeled i;

[0065] 3) When the length of the list is greater than 1, if list1 > list2, determine if list1 belongs to c in S. i If the data is of class i, then the data entry is labeled i;

[0066] 4) When the length of the list is greater than 1, if list1 = list2, determine whether list1 and list2 belong to the same category. If they do, determine whether list1 belongs to the c category in S. i If the data belongs to a class, then the data is marked with i; otherwise, the data is discarded.

[0067] 5) Input the obtained labeled dataset into the emoji prediction model for training.

[0068] Furthermore, step 6 includes the following steps:

[0069] Step 6-1: bpw calculation: This measures the embedding capacity of the hidden text's secret information; the larger the better. It is calculated using the following formula:

[0070]

[0071] Where bits represent the number of bits of secret information in the steganographic text, and words represent the number of words in the steganographic text;

[0072] Step 6-2: ΔMP Calculation: This measures how closely the stegtext and the real text are similar in average perplexity. The smaller the value, the better. Calculate using the following formula:

[0073] s = {x1, x2, ..., x n}

[0074]

[0075] ΔMP=|mean(PPL stego )-mean(PPL normal )|

[0076] Where s represents the sentence to be tested, n is a positive integer representing the sentence length, and in the perplexity calculation formula, 2≤j≤n, and j is a positive integer, mean(PPL) stego ) represents the average perplexity of all steganographic text, mean(PPL) normal This represents the average perplexity of all normal text.

[0077] Step 6-3: KLD (Kullback-Leibler divergence) measurement: The smaller the KLD, the more secure the steganography method. The calculation method is as follows:

[0078]

[0079] Where v x and v y Let μ and σ represent the vector representations of steganographic text and normal text in space, respectively; let σ and σ represent the mean and standard deviation of the sentence vectors, respectively.

[0080] Step 6-4: Use a steganalysis-based tool to test the steganalysis resistance of the three sets of steganographic texts. Evaluate the steganalysis resistance using commonly used metrics in classification tasks, namely accuracy (Acc) and recall (R). The concepts and formulas are described below:

[0081]

[0082]

[0083] Where TP represents the number of positive samples predicted as positive by the model, FP represents the number of negative samples predicted as positive, FN represents the number of positive samples predicted as negative, and TN represents the number of negative samples predicted as negative.

[0084] Step 6-5: Based on the experimental results, compare the security and anti-steganography capabilities of steganographic text with added emojis and plain text. At the same time, compare the advantages and disadvantages of the two recommendation strategies based on the index results.

[0085] The beneficial technical effects of this invention are as follows: Compared with previous methods, this invention overcomes the problem of lacking visual cues in steganographic text. Furthermore, based on the distribution of emojis in real-world environments, an emoji recommendation algorithm is designed to make the statistical distribution of emojis in steganographic text as close as possible to the statistical distribution of emojis in real-world environments. Ultimately, by incorporating visual cues, the security indicators of steganographic text are further improved. Attached Figure Description

[0086] Figure 1 This is a framework diagram of the deep learning-based method for enhancing the anti-steganography capabilities of steganographic text in this invention.

[0087] Figure 2 This is a diagram of the emoji insertion algorithm in this invention;

[0088] Figure 3 This is a diagram of the emoji replacement algorithm in this invention. Detailed Implementation

[0089] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the figures.

[0090] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0091] like Figure 1 As shown, this application proposes a deep learning-based method to enhance the steganalysis resistance of steganographic text. First, steganographic text is generated using a steganographic text generation model. Then, an emoji insertion algorithm is designed based on the relevant statistical characteristics of data from the real environment. Finally, an emoji prediction model is used to perform fine-grained sentiment analysis to insert emojis into the steganographic text, thereby improving its steganalysis resistance. Specifically, it includes the following steps:

[0092] Step 1: Collect a large amount of online text from social media platforms (any common social media platform). (In the following description, a single user post on a social media platform is referred to as online text.)

[0093] Step 1-1: First, use the common stop word library of the corresponding social media platform as the keyword library for the crawler.

[0094] Steps 1-2: Randomly select a stop word from the stop word library as the keyword for a round of data crawling. Since stop words exist in the online text collections of various topics, it is very appropriate to use stop words as the keywords for crawling general domain datasets.

[0095] Steps 1-3: You can customize the number of keywords crawled in each round, but it is recommended not to set it too high or too low. The goal is to ensure that all or most of the stop words in the stop word list have been used as keywords. This step aims to statistically analyze the characteristics of the communication channel itself, and the data volume should be no less than 1 million records.

[0096] Step 2: Preprocess the dataset accordingly. This includes removing special characters, hyperlinks, and phone numbers from the dataset.

[0097] Step 3: Analyze the statistical distribution characteristics of emojis in the data set and classify them according to semantics. Design two emoji recommendation algorithms (Insert and Replace, e.g.) Figure 2 and Figure 3 (As shown).

[0098] Step 3-1: Statistical analysis of emoji frequency characteristics: Statistical analysis of the frequency of each emoji and the proportion of online text containing emojis.

[0099] Step 3-2: Calculate the percentage of online text containing punctuation marks.

[0100] Step 3-3: Analyze the location information of emoticons in online text, dividing them into sentence beginning, sentence middle and sentence end.

[0101] Steps 3-4: Calculate the percentage of online text containing each number of emojis.

[0102] Steps 3-5: Based on the statistical information, design an emoji recommendation algorithm for macro-control.

[0103] Step 3-5-1: Algorithm Parameter Description:

[0104] A: The set of plain text steganography generated by the steganography generation model in step 5.

[0105] s: A single line of steganographic text in A.

[0106] s1: Steganographic text after the emoji is inserted.

[0107] B: The steganographic text set after adding emojis in step 5 using the emoji recommendation algorithm.

[0108] f: The percentage of online text containing emojis in the dataset obtained in step 3-1.

[0109] f1: The percentage of online text containing punctuation marks as calculated in step 3-2.

[0110] P: The location information of emoticons in online text obtained in step 3-3.

[0111] p: The specific value in P, which can be one of the following: the beginning of the sentence, the middle of the sentence, or the end of the sentence.

[0112] P1: A list of punctuation mark positions in the current web text.

[0113] p1: A specific value randomly selected from P1.

[0114] N: The percentage of online text containing each number of emojis as calculated in steps 3-4.

[0115] n: A specific value in N, with a range of n∈[1,5], where n is a positive integer.

[0116] Step 3-5-2: Design of the emoji insertion algorithm:

[0117] 6) Select s from A, and then determine whether to insert an emoji based on the probability value of f. If the result is False, proceed to 2); otherwise, proceed to 3).

[0118] 7) Output s to B, and the steganography operation ends.

[0119] 8) Select the quantity information n from N as the number of emojis to be inserted into s.

[0120] 9) The emoji prediction model performs sentiment prediction on the input sentence, then selects an emoji from the predicted emoji classification pool based on the true statistical distribution, and then selects the position information p from P as the insertion position of the emoji in s. This process is repeated n times.

[0121] 10) After completing operation 4), obtain s1 and output s1 to B.

[0122] Step 3-5-3: Design of the emoji replacement algorithm:

[0123] 9) Select s from A and determine whether s contains punctuation marks. If the result is True, proceed to 3); otherwise, proceed to 2).

[0124] 10) Output s to B, and the steganography operation ends.

[0125] 11) According to The probability value is used to determine whether to perform an emoji replacement operation. If the result is True, proceed to step 5); if the result is False, proceed to step 4).

[0126] 12) Output s to B, and the steganography operation ends.

[0127] 13) Select the quantity information n from N as the number of emojis to be inserted into s.

[0128] 14) Calculate the punctuation mark position index P1 of the current steganographic text, and randomly select a replacement position p1 from it.

[0129] 15) Use the emoji prediction model to make n predictions for s, and merge the n emojis into a string to replace the punctuation mark at position p1.

[0130] 16) After completing operation 7), obtain s1 and output s1 to B.

[0131] Step 4: Further process the dataset for training the stegtext generation model and the emoji prediction model. The stegtext generation model and the emoji prediction model are not limited to a specific model; they can be any deep learning model used for their respective tasks.

[0132] Step 4-1: Remove all emojis from the dataset to obtain a dataset in pure English text format for training the stegtext generation model.

[0133] Step 4-2: Extract all network text containing emojis from the dataset for training the emoji prediction model.

[0134] Step 4-2-1: Based on the statistical results in Step 3-1, construct an emoji library S. The construction standard should be to select emojis that account for more than 90% of the total number of emojis.

[0135] Step 4-2-2: Divide the emojis in the emoji library into m categories according to their semantics, and assign a number and Huffman coding to each category. The purpose of this is to embed certain secret information into the emojis. That is:

[0136] S = {c1, c2, ..., c m}

[0137] c i ={h1,h2,…,h n}

[0138] Where m represents the m categories of emojis in S according to their semantics, and c m c represents the m-th class; i Let h represent the i-th class (2≤i≤m) in S, where h n This represents the Huffman code of the nth emoji in this class.

[0139] Step 4-2-3: Construct data labels, the steps of which are as follows:

[0140] 6) Calculate the frequency of emoticons in a data point, sort them in descending order to form a list.

[0141] 7) When the length of the list is equal to 1, determine if list1 belongs to c in S. i If the data is of class i, then the data entry is labeled i.

[0142] 8) When the length of the list is greater than 1, if list1 > list2, determine if list1 belongs to listc in S. i If the data is of class i, then the data entry is labeled i.

[0143] 9) When the length of the list is greater than 1, if list1 = list2, determine whether list1 and list2 belong to the same category. If they do, determine whether list1 belongs to the c category in S. i If the data belongs to a class, then the data is marked with i; otherwise, the data is discarded.

[0144] 10) Input the obtained labeled dataset into the emoji prediction model for training.

[0145] Step 5: Use the stegtext generation model trained in Step 4 to generate stegtext with secret information, and input the stegtext into the emoji prediction model trained in Step 4. Add emojis to the stegtext according to the recommendation algorithm in Step 3.

[0146] Step 6: Calculate the basic indicators of the three sets of data generated in Step 5: the plain text steganography and the steganography with added emojis (the plain text steganography generated by the steganography generation model is one set, the steganography with added emojis via the Insert algorithm is another set, and the steganography with added emojis via the Replace algorithm is yet another set). Input these indicators into the steganography analysis model to compare the resistance to steganography analysis of the three sets of steganography.

[0147] Step 6-1: bpw (bits per word) calculation: This measures the embedding capacity of the hidden text's secret information; the larger the better. It is usually calculated using the following formula:

[0148]

[0149] Where bits represent the number of bits of secret information in the steganographic text, and words represent the number of words in the steganographic text.

[0150] Step 6-2: ΔMP Calculation: This measures how closely the average perplexity of the stegtext and the real text resembles each other. A smaller value is better. It is typically calculated using the following formula:

[0151] s = {x1, x2, ..., x n}

[0152]

[0153] ΔMP=|mean(PPL stego )-mean(PPL normal )|

[0154] Where s represents the sentence to be tested, and n is a positive integer representing the sentence length. In the perplexity calculation formula, 2 ≤ j ≤ n, and j is a positive integer. mean(PPL) stego The mean (PPL) represents the average perplexity of all steganographic texts in the experiment. normal () represents the average perplexity of all normal texts in the experiment.

[0155] Step 6-3: KLD (Kullback-Leibler divergence) measurement: KLD is an important metric proposed by Cachin to measure the security of steganography methods. It aims to measure the statistical distribution difference between stegtext and normal text. The smaller the value, the better, indicating that the steganography method is more secure. The calculation method is as follows:

[0156]

[0157] Where v x and v y Let μ and σ represent the vector representations of the steganographic text and the normal text in space, respectively. Let μ and σ represent the mean and standard deviation of the sentence vectors, respectively.

[0158] Step 6-4: Use steganalysis-based tools to test the steganalysis resistance of the three sets of steganographic texts. Evaluate the steganalysis resistance using several commonly used metrics in classification tasks, namely accuracy (Acc) and recall (R). The concepts and formulas are described below:

[0159]

[0160]

[0161] Among them, TP (true positive) represents the number of positive samples predicted as positive by the model, FP (false positive) represents the number of negative samples predicted as positive; FN (false negative) represents the number of positive samples predicted as negative, and TN (true negative) represents the number of negative samples predicted as negative.

[0162] Step 6-5: According to the specific experimental results, compare the security and steganalysis resistance of the stego text after adding emojis with that of the pure text stego text, and at the same time compare the advantages and disadvantages of the two recommended strategies according to the index results.

[0163] The applicant conducted experiments on a general domain dataset collected from common domestic and foreign social media platforms including Twitter and Weibo, and compared relevant indicators with the stego text generation model used (hereinafter collectively referred to as the base model). The larger the bpw, the better, and the smaller the other indicators, the better:

[0164] Regarding bpw: at an embedding rate of 1-3 bits, Insert > Replace > base model; at an embedding rate of 4 or higher, Replace < Insert < base model.

[0165] Regarding ΔMP: Insert < Replace < base model.

[0166] Regarding KLD: Insert < Replace < base model.

[0167] Regarding the Accuracy of the steganalysis model: Insert < Replace < base model.

[0168] Regarding the Recall of the steganalysis model: Insert < Replace < base model.

[0169] It should be noted that the feasibility of this method has been confirmed by experiments on common domestic and foreign social media platforms including Twitter, Weibo, etc. Therefore, this method is also applicable to other domestic and foreign social media platforms. In addition, for the two emoji recommendation strategies, after experimental comparison: the stego text under the Insert strategy has better indicator performance than the stego text under the Replace strategy. Therefore, it is recommended to adopt the Insert strategy for emoji recommendation operations on stego text.

[0170] The above description is merely a preferred embodiment of the present invention. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the technical solutions of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention are all within the scope of protection of the technical solutions of the present invention.

Claims

1. A method for enhancing the anti-steganography capabilities of steganographic text based on deep learning, characterized in that, Includes the following steps: Step 1: Collect online text from social media platforms; Step 2: Perform appropriate preprocessing on the dataset, including removing special characters, hyperlinks, and phone numbers to form the dataset; Step 3: Collect statistical distribution features of emojis in the data set and classify them according to semantics, and design two emoji recommendation algorithms: Insert algorithm and Replace algorithm; Algorithm parameter description: A: The set of plain text steganography generated by the steganography generation model in step 5; s: A single line of steganographic text in A; s1: Steganographic text after the emoji is inserted; B: The steganographic text set after adding emojis in step 5 using the emoji recommendation algorithm; f: The percentage of online text containing emojis in the dataset obtained in step 3-1; f1: The percentage of online text containing punctuation marks as calculated in step 3-2; P: Location information of emoticons in online text obtained in step 3-3; p: The specific value in P, which can be one of the following: the beginning of the sentence, the middle of the sentence, or the end of the sentence; P1: A list of punctuation mark positions in the current web text; p1: A specific value randomly selected from P1; N: The percentage of online text containing each number of emojis as calculated in steps 3-4; n: A specific value in N, with a range of n∈[1,5], where n is a positive integer; The Insert algorithm includes the following steps: 1) Select s from A, and then determine whether to insert an emoji based on the probability value of f. If the result is False, proceed to 2); if it is True, proceed to 3). 2) Output s to B, and the steganography operation ends; 3) Select the quantity information n from N as the number of emojis to be inserted into s; 4) The emoji prediction model performs sentiment prediction on the input sentence, then selects an emoji from the predicted emoji classification pool based on the true statistical distribution, and then selects the position information p from P as the insertion position of the emoji in s. This process is repeated n times. 5) After completing operation 4), obtain s1 and output s1 to B; The Replace algorithm includes the following steps: 1) Select s from A and determine whether s contains punctuation marks. If the result is True, proceed to 3); otherwise, proceed to 2). 2) Output s to B, and the steganography operation ends; 3) According to The probability value is used to determine whether to perform an emoji replacement operation. If the result is True, proceed to step 5); if the result is False, proceed to step 4. 4) Output s to B, and the steganography operation ends; 5) Select the quantity information n from N as the number of emojis to be inserted into s; 6) Calculate the punctuation mark position index P1 of the current steganographic text, and randomly select a replacement position p1 from it; 7) Use the emoji prediction model to make n predictions for s, and merge the n emojis into a string to replace the punctuation mark at position p1; After completing operation 7), s1 is obtained and output to B; Step 4: Further process the dataset for training the stegtext generation model and the emoji prediction model respectively; Step 5: Use the stegtext generation model trained in Step 4 to generate stegtext with secret information, and input the stegtext into the emoji prediction model trained in Step 4. Add emojis to the stegtext according to the recommendation algorithm in Step 3. Step 6: Calculate the basic indicators of the three sets of data: the plain text steganography generated by the steganography generation model in Step 5, the steganography with emojis added by the Insert algorithm, and the steganography with emojis added by the Replace algorithm. Input these indicators into the steganography analysis model to compare the resistance to steganography analysis of the three sets of steganography.

2. The method for enhancing the anti-steganography capability of steganographic text based on deep learning according to claim 1, characterized in that, Step 1 includes the following steps: Step 1-1: First, use the general stop word list of the corresponding social media platform as the keyword library for the web crawler; Step 1-2: Randomly select a stop word from the stop word library as the keyword for one round of data crawling; Steps 1-3: Customize the number of keywords crawled in each round, aiming to statistically analyze the characteristics of the communication channel itself, with a data volume of no less than 1 million entries.

3. The method for enhancing the anti-steganography capability of steganographic text based on deep learning according to claim 1, characterized in that, Step 3 includes the following steps: Step 3-1: Statistical analysis of emoji frequency characteristics: Statistical analysis of the frequency of each emoji and the percentage of online text containing emojis; Step 3-2: Calculate the percentage of online text containing punctuation marks; Step 3-3: Analyze the location information of emoticons in online text, dividing them into sentence beginning, sentence middle, and sentence end; Steps 3-4: Calculate the percentage of online text containing each number of emojis; Steps 3-5: Based on the statistical information, design an emoji recommendation algorithm for macro-control.

4. The method for enhancing the anti-steganography analysis capability of steganographic text based on deep learning according to claim 3, characterized in that, Step 4 includes the following steps: Step 4-1: Remove all emojis from the dataset to obtain a dataset in pure English text format for training the stegtext generation model; Step 4-2: Extract all network text containing emojis from the dataset for training the emoji prediction model.

5. The method for enhancing the anti-steganography capability of steganographic text based on deep learning according to claim 4, characterized in that, Step 4-2 includes the following steps: Step 4-2-1: Based on the statistical results in Step 3-1, construct an emoji library S. The construction criterion is to select emojis that account for more than 90% of the total number of emojis. Step 4-2-2: Divide the emojis in the emoji library into m categories according to their semantics, and assign a number and Huffman coding to each category, i.e.: S={c1,c2,…,c m } c i ={h1,h2,…,h n } Where m represents the m categories of emojis in S according to their semantics, and c m c represents the m-th class; i Let h represent the i-th class in S, where 2≤i≤m, and h n This represents the Huffman code of the nth emoji in this class; Step 4-2-3: Construct data labels.

6. The method for enhancing the anti-steganography capability of steganographic text based on deep learning according to claim 5, characterized in that, Step 4-2-3 includes the following steps: 1) Calculate the frequency of emoticons in a data point, sort them in descending order to form a list; 2) When the length of the list is equal to 1, determine if list1 belongs to c in S. i If the data is of class i, then the data entry is labeled i; 3) When the length of the list is greater than 1, if list1 > list2, determine if list1 belongs to c in S. i If the data is of class i, then the data entry is labeled i; 4) When the length of the list is greater than 1, if list1 = list2, determine whether list1 and list2 belong to the same category. If they do, determine whether list1 belongs to the c category in S. i If the data belongs to a class, then the data is marked with i; otherwise, the data is discarded. 5) Input the obtained labeled dataset into the emoji prediction model for training.

7. The method for enhancing the anti-steganography capability of steganographic text based on deep learning according to claim 1, characterized in that, Step 6 includes the following steps: Step 6-1: bpw calculation: This measures the embedding capacity of the hidden text's secret information; the larger the better. It is calculated using the following formula: Where bits represent the number of bits of secret information in the steganographic text, and words represent the number of words in the steganographic text; Step 6-2: ΔMP Calculation: This measures how closely the stegtext and the real text are similar in average perplexity. The smaller the value, the better. Calculate using the following formula: s={x1,x2,…,x n } ΔMP=|mean(PPL stego )-mean(PPL normal )| Where s represents the steganographic text to be tested, n is a positive integer representing the sentence length, and in the perplexity calculation formula, 2≤j≤n, and j is a positive integer, mean(PPL) stego This represents the average perplexity of all steganographic text. mean(PPL normal This represents the average perplexity of all normal text. Step 6-3: KLD (Kullback-Leibler divergence) measurement: The smaller the KLD, the more secure the steganography method. The calculation method is as follows: Where v x and v y Let μ and σ represent the vector representations of steganographic text and normal text in space, respectively; let σ and σ represent the mean and standard deviation of the sentence vectors, respectively. Step 6-4: Use a steganalysis-based tool to test the steganalysis resistance of the three sets of steganographic texts. Evaluate the steganalysis resistance using commonly used metrics in classification tasks, namely accuracy (Acc) and recall (R). The concepts and formulas are described below: Where TP represents the number of positive samples predicted as positive by the model, FP represents the number of negative samples predicted as positive, FN represents the number of positive samples predicted as negative, and TN represents the number of negative samples predicted as negative. Step 6-5: Based on the experimental results, compare the security and anti-steganography capabilities of steganographic text with added emojis and plain text. At the same time, compare the advantages and disadvantages of the two recommendation strategies based on the index results.

Images