A sequence-generating dna steganography method and evaluation method

Abstract

Existing steganography techniques generate pseudo-sequences that retain only a small portion of the statistical distribution features of the original DNA sequence, resulting in poor concealment and weak resistance to steganalysis, thus leading to low security. This invention proposes a sequence-generating DNA steganography method based on conditional probability adaptive coding. First, an RNN network is used to learn the statistical distribution features of natural DNA. Then, an adaptive grouping algorithm generates pseudo-sequences highly similar to the original DNA. Experimental results show that this method significantly improves the concealment and resistance to steganalysis of the pseudo-sequences, enhancing the security of information hiding. Furthermore, given the lack of a comprehensive evaluation mechanism for the concealment of pseudo-sequences generated by existing DNA steganography techniques, this invention proposes using three parameters to evaluate the perceptual and statistical concealment of the pseudo-sequences, further refining the evaluation system for the security of DNA steganography systems.

Description

Technical Field

[0001] This invention relates to the field of encrypted information transmission, specifically to a secret information transmission technology based on DNA sequence encoding technology, and more specifically to a sequence-generating DNA steganography method and evaluation method, including: a sequence-generating DNA steganography method, a method for obtaining a steganalytic vector pseudo-sequence (hereinafter referred to as pseudo-sequence), and a method for evaluating it. Background Technology

[0002] In today's information transmission environment, the security of information transmission has always been a major concern. Steganography is one of the important technologies to ensure information security. It can hide secret information in natural carriers without arousing suspicion from third-party inspectors, thereby ensuring the secure transmission of secret messages.

[0003] In modern steganography, the form of the natural medium can be divided into digital and physical forms. Digital mediums offer faster transmission rates and more diverse signal formats (images, videos, audio, etc.), but they are heavily reliant on the internet or communication devices. Furthermore, network noise and environmental interference can negatively impact the robustness of steganography algorithms. Therefore, in certain specific scenarios (limited network access, severe interference, etc.), physical mediums play an irreplaceable role, making physical medium-based steganography a subject of ongoing interest for researchers in the field.

[0004] With the rapid development of gene editing and deoxyribonucleic acid (DNA) synthesis technologies in recent years, DNA-based steganography has become a major focus in physical steganography. Using DNA as a steganographic carrier offers three advantages: First, it has an extremely large information hiding capacity; one gram of DNA can hide 10... 6 The data is in Tb. Secondly, due to the high diversity and complexity of DNA, and its tiny physical scale, conventional steganalysis methods struggle to detect statistical distortions in the stegographic carrier after information is hidden, and these distortions are not readily perceptible to humans. Thirdly, current mature DNA synthesis technology offers low synthesis costs. Combining these advantages, DNA has become a new generation of molecular-level stegographic carrier with high hiding capacity, high imperceptibility, and high feasibility.

[0005] In natural DNA, regions can be divided into coding and non-coding regions based on whether their base sequences express proteins. Hiding information in coding regions requires maintaining the biological characteristics of the base sequences before and after information hiding. A common information hiding strategy is to replace wobbly bases using codon degeneracy rules. However, coding regions have numerous biological limitations, often providing only a very low hiding capacity. In contrast, non-coding regions do not express proteins, and each base in these regions can hide information; therefore, non-coding DNA can provide a much higher hiding capacity than coding DNA. Considering these factors, non-coding DNA is increasingly being used in DNA steganography.

[0006] In general, DNA steganography algorithms can be divided into two types: substitution steganography and generative steganography. Among them, substitution steganography generates pseudo sequences by replacing the bases of the original DNA sequence or encoding the secret information into base fragments and then inserting them into the original DNA sequence. During decoding, the sender and receiver share the original vector, compare it with the pseudo sequence, detect the replaced bases, and complete the extraction of secret information. In the 2010 paper [1], Shiu, Hung-Jr et al. proposed the random insertion method and the complementary pair method. In the 2012 paper [2], Jin-Shiuh Taur et al. proposed the lookup table substitution method (TLSM), and in the same year, in the paper [3], C. Guo proposed the repeated base substitution method (RBS). In 2013 and 2016, in the papers [4][5], Amal Khalifa et al. proposed the least significant bit (LSB) synonymous base substitution method and the universal complementary base substitution method (GCBS), respectively. However, these substitutions all result in significant statistical distortion in the pseudo-sequences compared to the natural DNA sequence.

[0007] Generative steganography directly generates pseudo-sequences through specific generative constraints, such as constraints on the codon content and information content of the original carrier. Compared to substitution steganography, generative steganography typically has lower distortion and the sender and receiver do not need to share the original carrier, making it more feasible in real-world scenarios. Although generative steganography has many advantages over substitution steganography, in practical applications, generative steganography algorithms need to learn and maintain numerous statistical features of the natural DNA sequence to ensure that they are not distorted after information hiding. That is, the generated pseudo-sequence must maintain consistency with the statistical characteristics of the natural sequence as much as possible to ensure strong concealment and resistance to detection by various steganographic analysis techniques. Therefore, generative steganography is more difficult to implement than substitution steganography. Currently, mainstream generative steganography techniques targeting non-coding regions of DNA lack the ability to model the statistical characteristics of non-coding DNA regions, resulting in severe distortion after pseudo-sequence synthesis, leading to low concealment and security of the pseudo-sequences. For example, the earliest DNA-generating steganography can be traced back to 1999. In paper [6], CTC Clelland et al. directly generated pseudo sequences from the content of secret information. In 2003, in paper [7], GCSmith et al. proposed comma encoding and alternation encoding to generate pseudo sequences. In 2015, in paper [8], Ji Young Chun et al. generated pseudo sequences based on the codon distribution of the original DNA sequence. None of these methods can avoid serious statistical distortion in pseudo sequences, resulting in low security of pseudo sequences.

[0008] In addition, the security of pseudo-sequences can also be reflected by the anti-steganography capability of the vector. Steganography is a technique that determines whether a data contains secret information based on observed data. Steganography in coding region DNA can effectively detect whether it contains secret information by using synonymous base matching. For non-coding region DNA, dinucleotide relative abundance spectrum anomaly analysis is often used. Currently, steganalysis models based on machine learning and deep learning, such as the latest random forest model, the RNN+CNN steganalysis model proposed by Ho bae in 2020 in paper [9], and the TEXTCNN model proposed by Chen, Yahui et al. in paper

[15] , have detection accuracy far exceeding that of traditional steganalysis models. Furthermore, due to the excellent feature extraction and analysis capabilities of neural networks, current steganalysis technology has seriously threatened the security of steganalysis algorithms.

[0009] In summary, existing steganography techniques generate pseudo-sequences that retain only a small portion of the statistical distribution characteristics of the original DNA sequence, resulting in poor concealment. Furthermore, under current steganalysis detection models, their detection accuracy is close to 100%, and their resistance to steganalysis is also very poor, thus offering low security. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention proposes a sequence-generating DNA steganography method based on conditional probability adaptive coding. First, an RNN network is used to learn the statistical distribution characteristics of natural DNA. Then, an adaptive grouping algorithm generates pseudo-sequences highly similar to the original DNA. Experimental results show that this method significantly improves the concealment and resistance to steganalysis of the pseudo-sequences, thereby enhancing the security of information hiding. Furthermore, given the lack of a robust evaluation mechanism for the concealment of generated pseudo-sequences in existing DNA steganography techniques, this invention proposes using three parameters: GC content deviation (GC...). d Melting temperature deviation (Tm) d KL divergence (KL) of relative abundance of dinucleotides ρ *) to evaluate the perceptual and statistical concealment of pseudo-sequences, further improving the evaluation system for the security of DNA steganography systems.

[0011] To achieve the above objectives, the present invention adopts the following technical solution, including:

[0012] (I) This invention provides a generative steganography method based on conditional probability adaptive coding, comprising the following steps:

[0013] Step 1: Select the DNA sequence to be used for RNN network learning. Since the non-coding region of DNA does not express genetic information, this region has fewer restrictions and a larger hiding capacity, making it more suitable for DNA steganography. Therefore, the DNA sequences selected in this invention are all non-coding DNA sequences.

[0014] Step 2: Preprocess the DNA sequence. Divide the DNA sequence into segments of different lengths, denoted as SL (SL∈{2, 3, 4, 5, 6, 8, 10, 20, 33, 40, 50, 66}). The base sequence is preserved during segmentation. Each SL base is considered a base unit after segmentation, and a base unit dictionary (V) corresponding to the DNA sequence is established. b ).

[0015] Step 3: Obtain the dataset corresponding to the DNA sequence. According to the arrangement order of the divided base units, several base units are combined into several short base chains of equal length. Since there is a small error in single-step base synthesis, according to the suggestion of J. Bornholt et al. in paper

[10] , the length of the synthesized DNA should not exceed 200. Therefore, based on the parity of SL, the present invention sets the length of a single short base chain to the maximum value within 200 (for example, if SL is 3 or SL is 4, the length of a single short base chain is 198 or 200). According to the above measures, the dataset corresponding to the DNA sequence is obtained.

[0016] As an alternative implementation, the base units are preprocessed and then word-embedded, mapping each base unit to a high-dimensional word vector.

[0017] As an alternative implementation, the word vector dimension is set to 350.

[0018] Step 4: Construct an RNN neural network to learn the statistical distribution characteristics of the above dataset.

[0019] As an alternative implementation, the RNN neural network uses Long Short-Term Memory (LSTM) units, which contain two hidden layers, each with 512 LSTM units. To prevent overfitting, Srivastava et al.

[11] proposed the concept of dropout rate, which represents the proportion of neurons that are randomly deactivated each time the parameters are updated. In this patent, the dropout rate is set to 0.2.

[0020] As an alternative implementation, multiple high-dimensional vectors can be stacked as input to a single RNN network to accelerate the model training process.

[0021] As an alternative implementation, the number of stacked vectors is 30.

[0022] Step 5: Input the state of the last hidden layer of the neural network into the fully connected layer. The input parameter dimension of the fully connected layer is the number of neurons in a single hidden layer, and the input parameter dimension is V. b The number of base units in it.

[0023] After obtaining the output of the fully connected layer, it needs to be normalized to a probability distribution. As an optional implementation method, the Softmax method proposed by John D. Cook et al. in paper

[12] is selected to obtain V. b Probability distribution of all base units in (P) t ).

[0024] Construct a loss function to determine whether to continue training, and decide whether to continue adjusting the parameter values ​​of the neural network based on whether the loss function converges.

[0025] As an alternative implementation method, the loss function for determining whether to continue training is:

[0026]

[0027] Where x' i It is a tag of a base unit, x i is the probability value corresponding to the base unit, and k is the number of base units in a single short chain.

[0028] Step 6: Place P t The input is fed into the loss function to determine whether to continue adjusting the RNN network parameters. Once the loss function converges, the trained steganalytic vector pseudo-sequence generation model is obtained.

[0029] As an alternative implementation, the present invention uses the AdamW optimizer and backpropagation algorithm in paper

[13] to correct the RNN network parameters.

[0030] As an alternative implementation, the learning rate of the present invention is 0.0001, and the number of training epochs is 150.

[0031] During the information embedding stage, the generative model returns a probability set of all base units in the base unit dictionary based on the generated steganalytic vector pseudo-sequence, selects different encoding algorithms to encode the probability set, and generates a pseudo-sequence containing secret information.

[0032] As an alternative implementation, a random binary bit stream is used as the secret information.

[0033] As an alternative implementation, the encoding algorithm selected is the Adaptive Dynamic Grouping (ADG) algorithm, which encodes the probability set returned by the generative model and automatically generates a pseudo sequence containing secret information based on the secret information to be hidden.

[0034] (II) This invention provides a method for selecting the optimal steganalytic vector pseudo-sequence under different segmentation lengths (SL), comprising the following steps:

[0035] Step 1: Calculate the binary classification accuracy (Accuracy, Acc) of each pseudo-sequence under SL and the original sequence, and calculate the embedding rate of each pseudo-sequence under SL. The embedding rate is expressed as Bits Per Nucleotide (BPN).

[0036] The formula for calculating Acc is as follows: The pseudo-sequence and the original sequence are designated as positive and negative samples, respectively. TP and TN represent the number of correctly judged samples in the positive and negative samples, respectively, and FP and FN represent the number of incorrectly judged samples C in the positive and negative samples, respectively.

[0037] The formula for calculating BPN is as follows: Where L MSG With L B These represent the number of bits of the secret information in binary and the total number of bases in the pseudo-sequence, respectively.

[0038] Step 2: Since different SLs can affect the security and capacity of pseudo-sequences, an excessively large SL can lead to feature sparsity and make the base unit dictionary unable to meet the size of the dictionary matched by the SL (for example, if the SL is 8, the size of its matched base unit dictionary should be 4). 8 This leads to an increase in Acc and a decrease in BPN. Similarly, a small SL results in insufficient features, making it impossible to learn the statistical characteristics of the original DNA sequence. The pseudo-sequences obtained in both cases are suboptimal. To find the optimal pseudo-sequence under different SLs, this invention proposes the Effective Bits Per Nucleotide (EBPN) for determination.

[0039] The formula for calculating EBPN is: EBPN=2*(1-Acc)*BPN.

[0040] Step 3: By calculating the EBPN value, when the EBPN of the pseudo-sequence corresponding to a certain SL is the largest, the pseudo-sequence under this SL can be determined as the best stegtext carrier pseudo-sequence.

[0041] (III) Addressing the lack of a robust evaluation mechanism for the concealment of generated DNA steganalyte pseudo-sequences in existing methods, this invention proposes three indicators reflecting the perceived and statistical concealment of DNA pseudo-sequences, including: GC content deviation (GC... d ), melting temperature deviation (Tm) d KL divergence (KL) of relative abundance of dinucleotides ρ *). This method evaluates the perceptual and statistical concealment of pseudo-sequences in steganographic carriers, further improving the security indicators of steganography technology.

[0042] In DNA sequences, the GC content parameter refers to the proportion of total G and C bases to all bases, while the melting temperature parameter refers to the temperature at which the DNA double helix unwinds halfway. These two parameters are commonly used in DNA analysis. Although DNA molecular-level carriers are highly concealed from human perception, differences between different molecular-level carriers can still be detected by analyzing the differences in these two parameters.

[0043] Therefore, to further evaluate the perceptual concealment of pseudo-sequences, this invention proposes adding two new parameters: GC content bias (GC content bias). d ) and melting temperature deviation (Tm) d (This is used to measure the perceptual concealment of pseudo-sequences.)

[0044] Among them, GC content deviation (GC) d The formula for calculating ) is:

[0045]

[0046] Among them, GC r GC f These represent the GC content in the original DNA sequence and the pseudo-sequence of the stegtext vector, respectively.

[0047] Melting temperature deviation (Tm) d The formula for calculating ) is:

[0048]

[0049] Where Tm r Tm f These represent the melting temperatures of the original DNA sequence and the stegtext vector pseudo-sequence, respectively.

[0050] In addition, in paper

[14] , MBBeck et al. pointed out in 2014 that statistical analysis of abnormal expression of dinucleotide relative abundance profiles can provide support for the detection of statistical concealment of DNA steganalysis, but no specific scheme was provided in the paper. In order to accurately reflect the statistical concealment of pseudo-sequences, this invention calculates another new index parameter: the KL divergence of dinucleotide relative abundance (KL ρ *) is used to reflect the differences in the distribution of dinucleotides in the vector, and its expression is:

[0051]

[0052] in and These represent the relative abundance of different dinucleotides X and Y among the sixteen dinucleotides in the original DNA sequence and pseudo-sequence, respectively.

[0053] The beneficial effects of this invention are:

[0054] The method of this invention first ensures that the generated DNA sequence, after hiding information, can still maintain the statistical characteristics of natural DNA without distortion, i.e., high concealment. Secondly, it must ensure that the steganalytic vector pseudo-sequence generated by this steganography algorithm has good resistance to steganalysis techniques, i.e., strong anti-steganography capability.

[0055] This invention proposes a sequence-generating DNA steganography method based on conditional probabilistic adaptive coding (ADG). It utilizes an RNN network to learn the statistical distribution characteristics of the original DNA, and then uses an adaptive grouping algorithm to generate pseudo-sequences highly similar to the original DNA. The generated pseudo-sequences possess high perceptual and statistical concealment. Simultaneously, this method significantly improves the resistance to steganalysis of the pseudo-sequences, making the information hiding of the pseudo-sequences highly secure. Experimental analysis and verification using the pseudo-sequences generated by the proposed sequence-generating DNA steganography method show that the generated pseudo-sequences exhibit good perceptual and statistical concealment. Furthermore, their resistance to steganalysis is significantly superior to current DNA steganography methods.

[0056] Furthermore, since existing DNA steganography lacks a comprehensive evaluation mechanism for the concealment of the generated DNA pseudo-sequences, this patent proposes an evaluation mechanism that uses three parameters as evaluation criteria: KL divergence (KLρ*) of dinucleotide relative abundance, GC content deviation (GCd), and melting temperature deviation (Tmd). This evaluation mechanism can reflect the statistical concealment and perceived concealment of DNA pseudo-sequences at both the statistical and physical property levels, further improving the security evaluation mechanism in this field and enhancing the security of DNA pseudo-sequences.

[0057] The advantages of this invention are specifically manifested in the following aspects:

[0058] This invention provides a sequence-generating DNA steganography method based on conditional probability adaptive coding (ADG), which, in the DNA sequence preprocessing step, provides a method for selecting the optimal steganalytic vector pseudo-sequence under different segment lengths SL.

[0059] Secondly, the stegographic vector pseudo-sequence generated by the method described in this invention largely retains the statistical characteristics of the original sequence, making it difficult to be detected by the latest DNA steganalysis models. It has excellent concealment and strong resistance to steganalysis, effectively improving the security of DNA pseudo-sequences.

[0060] Furthermore, this invention proposes new indicators for evaluating the perceptual and statistical concealment of DNA pseudo-sequences, further improving the security evaluation mechanism for DNA steganography technology.

[0061] Experimental analysis and verification show that the pseudo-sequence generated by this invention has better concealment and stronger resistance to steganalysis compared with pseudo-sequences generated by other DNA steganography techniques.

[0062] Furthermore, this invention has full compatibility with binary codes, and the proposed method can be combined with any DNA cryptography to further improve the security of covert communication. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the algorithm for embedding secret information in this invention.

[0064] Figure 2 This is a schematic diagram of the algorithm for extracting secret information in this invention.

[0065] Figure 3 This is the overall flowchart of the present invention. Box ① represents model training and learning, and box ② represents the overall information hiding process.

[0066] Figure 4 A schematic diagram for obtaining the probability set of all base units.

[0067] Figure 5 A detailed diagram illustrating information hiding.

[0068] Figure 6 This is a statistical graph of single base bits (BPNs) for the control method.

[0069] Figure 7 This figure shows the results of the anti-steganography analysis capabilities of the pseudo-sequences of the stegographic vector generated by the control method and the present invention. The subscripts represent the standard deviations of multiple repeated experiments; similar subscripts all represent the standard deviations of repeated experiments.

[0070] Figure 8 This is a dimensionality reduction visualization of the statistical characteristics of the pseudo-sequences of the stegtext vector generated by the control method and the present invention.

[0071] Figure 9 The diagram shows the concealment analysis results of the pseudo-sequences of the stegographic vector generated by the control method and the present invention.

[0072] Figure 10 The diagram shows the effective single-base bit capacity of the pseudo-sequence of the stegtext vector generated by the control method and the present invention. Specific implementation methods

[0073] The technical features of the present invention will be described in detail with reference to the accompanying drawings.

[0074] A sequence-generating DNA steganography method, based on conditional probabilistic adaptive coding (ADG), utilizes an RNN network to learn the statistical distribution characteristics of the original DNA, and then uses an adaptive grouping algorithm to generate a pseudo-sequence of the steganalytic vector that is highly similar to the natural DNA and contains secret information. This pseudo-sequence is then synthesized using biosynthesis techniques according to the pseudo-sequence to obtain a DNA strand containing the secret information. This DNA strand exhibits high perceptual and statistical concealment, significantly improving the resistance to steganalysis and ensuring high security in information hiding.

[0075] It is important to note that different steganalysis methods are required for different steganalytic vectors. Natural DNA strands can be divided into coding and non-coding regions. Coding regions express proteins, and not all bases in these regions can be used to hide information; arbitrarily changing bases will alter the biological characteristics expressed by the DNA strand and the function of the protein. Non-coding regions, on the other hand, do not express proteins, and every base can hide information, thus possessing a higher hiding capacity. Therefore, the focus of steganalysis research on these two regions differs. For coding regions, the emphasis should be on whether the sequence transcribes the protein intact after steganalysis, while for non-coding regions, the emphasis should be on whether numerous statistical and physical properties (such as GC ratio) of the sequence undergo significant changes after steganalysis. Therefore, for different steganalytic vectors, selecting an appropriate steganalysis method is crucial for successful implementation.

[0076] Furthermore, the sequence-generating DNA steganography method of the present invention includes the following steps:

[0077] Step 1: Preparation phase: Select the DNA sequence used for RNN network learning, which is the non-coding region of a natural DNA sequence.

[0078] Step 2: Preprocessing stage: The selected DNA sequence is divided into segments of different lengths (SL) for preprocessing to obtain a base dictionary V corresponding to the DNA sequence. b The value of the segment length is called the SL value.

[0079] Step 3: Compile the preprocessed DNA sequences into a dataset.

[0080] Step 4: Training Phase of the Stegographic Vector Pseudo-Sequence Model: Construct the hidden layers and fully connected layers of the RNN neural network. The RNN neural network is used to learn the statistical distribution characteristics of the dataset obtained in Step 3. The hidden layers of the RNN neural network save the learned statistical distribution by updating the neuron weights and generate a pseudo-sequence model. The fully connected layers of the RNN neural network act as the receivers of the RNN network output, receiving the RNN network output and finally outputting the base unit dictionary V. b The fractions of all base units in the set are normalized to obtain the probability set P. t , used to calculate the loss function;

[0081] Step 5: Construct the loss function. Whether training continues depends on whether the loss function converges: Adjust the neuron weight parameters of the RNN hidden layer; that is, adjust the base unit probability distribution P obtained in Step 4. tThe parameters are input into the loss function, and whether the loss function converges determines whether to continue adjusting the neural network parameters. Once the loss function converges, the trained steganalytic vector pseudo-sequence generation model is obtained, and the process proceeds to step 6. Otherwise, step 5 is repeated until the loss function converges.

[0082] Step 6: Secret Information Embedding Stage: Construct a pseudo-sequence containing secret information based on the generative model obtained in Step 5. Specifically, based on the base unit dictionary V obtained in Step 2... b The probability set of all base units in the DNA is used to encode the probability set of base units using the ADG coding algorithm to generate a pseudo-sequence containing secret information. This pseudo-sequence containing secret information is then used to synthesize a steganographic DNA strand using biosynthesis technology, and can then be sent to the recipient.

[0083] See Figure 1 Furthermore, the decryption method, or the revealing step, can be considered the reverse of the encryption process, and the key is the same, i.e., a symmetric encryption algorithm. The aforementioned key is the DNA steganalytic vector pseudo-sequence. The revealing method is as follows:

[0084] Step C1: Steganographic DNA Strand Transmission Stage: The sender transmits the stegographic DNA strand to the receiver using any method suitable for transmitting natural DNA strands. As an optional implementation, it can be dissolved in pure water before being sent to the receiver. Simultaneously, to facilitate the receiver's acquisition of the secret information, the sender must also send the receiver the DNA strand synthesized with the pseudo-sequence containing the secret information from step 7, the training model obtained in step 5, the SL value obtained in step 2, and the base unit dictionary V. b .

[0085] Step C2: Secret Information Extraction Stage: The information extraction and information hiding processes are reversed. Therefore, after receiving the steganalyzed DNA strand, the receiver uses biotechnology to reverse-engineer and analyze it, obtaining the sequence of the DNA strand, i.e., the pseudo-sequence containing the secret information. The pseudo-sequence is then segmented according to the SL value to obtain the processed sequence composed of base units. A base unit is the number of bases, SL.

[0086] Step C3: Input the sequence obtained in step C2 into the received training model to obtain V. b The probability distribution P of all base units in t .

[0087] Step C4: Use the ADG encoding algorithm to process P tEncode the sequence and extract the secret information contained in the last word of the base unit sequence. After extraction, delete the last base unit, update the base unit sequence, and repeat steps C3 and C4 until all information in all base units has been extracted. At this point, the receiver has obtained all the secret information.

[0088] Furthermore, the steganography method described in this patent is actually a computer algorithm, therefore, it requires, but is not limited to, a computer to execute. In nature, DNA chains are chain-like structures formed by the four bases ATGC arranged in different sequences and lengths. In this patent, a computer neural network is used to learn the arrangement rules and statistical distribution characteristics of natural DNA chains, and then insert secret information to generate a DNA stegographic vector pseudo-sequence that is extremely similar to natural DNA. This newly generated sequence is essentially a string of letters composed of ATGC, not a physical object. This sequence is then sent to a biotechnology company, which can generate a physical DNA chain containing the secret information. After receiving the physical DNA chain, the recipient can reverse-order it using biological processing methods to obtain the ATGC sequence in the DNA chain, which is this string of letters (pseudo-sequence). Then, the hidden secret information is obtained through the decryption and extraction method described in the patent. Specifically, the biotechnology company synthesizes DNA according to the sequence and reverse-orders the obtained DNA.

[0089] Furthermore, in step 1 of the sequence-generating DNA steganography method described in this invention, the DNA sequences selected for learning by the RNN network are all non-coding DNA sequences.

[0090] Furthermore, in step 2, the specific method for preprocessing the DNA sequence is as follows: the DNA sequence is divided into segments of different lengths, denoted as SL. SL∈{2, 3, 4, 5, 6, 8, 10, 20, 33, 40, 50, 66}. The base sequence is preserved during segmentation. After segmentation, each SL base is considered a base unit, and a base unit dictionary V corresponding to the DNA sequence is established. b .

[0091] Furthermore, the dataset corresponding to the DNA sequence in step 3 is obtained as follows: based on the arrangement order of the base units after partitioning in step 2, the base units are combined into two or more short base strands. These short base strands are of equal length, and the maximum length of a single short base strand is less than 200. This yields the dataset corresponding to the DNA sequence. As an optional implementation, word embedding is performed on the base units after preprocessing, mapping each base unit to a high-dimensional word vector. As an optional implementation, the word vector dimension is set to 350 dimensions.

[0092] Furthermore, in step 4, the RNN neural network employs Long Short-Term Memory (LSTM) units, which consist of two hidden layers, each with 512 LSTM units. Alternatively, high-dimensional vector stacks can be used as input to the RNN network in a single pass to accelerate the model training process. The number of stacked vectors is 30.

[0093] Furthermore, in step 5: the RNN neural network has multiple hidden layers, each composed of multiple neurons. First, the state of the last hidden layer is input to the fully connected layer. The fully connected layer is a single-layer neural network, composed of several neurons. Any neuron can receive all the outputs of the hidden layers and output its own score, ultimately obtaining the set of scores from all neurons in that layer. The dimension of the input vector of the fully connected layer is the number of neurons in the last hidden layer. Since it is necessary to obtain the scores of all base units in the base dictionary, the number of neurons in the fully connected layer is equal to the base dictionary V. b The number of base units in the vector, i.e., the dimension of the output score vector, is V. b The number of base units. The dimension of the input parameters of the connection layer, that is, how dimensional the input vector is, representing the number of neurons in the last hidden layer. After obtaining the set of scores output by the fully connected layer, it needs to be normalized to a probability distribution. Then, the Softmax method is used to obtain V. b Probability distribution of all base units in (P) t ).

[0094] Furthermore, in step 5, the loss function formula is:

[0095] Where, x' i It is a tag of a base unit, x i is the probability value corresponding to the base unit, and k is the number of base units in a single short chain.

[0096] Furthermore, in step 5: the AdamW optimizer and backpropagation algorithm are used to correct the RNN network parameters. Preferably, the learning rate is 0.0001 and the number of training epochs is 150.

[0097] Furthermore, in step 6: a random binary bit stream is used as the secret information. The encoding algorithm chosen is the Adaptive Dynamic Grouping (ADG) algorithm, which encodes the probability set returned by the generative model and automatically generates a pseudo-sequence containing the secret information according to the secret information to be hidden.

[0098] The pseudo-sequence acquisition method of the sequence generation DNA steganography method described in this invention is obtained through the following steps:

[0099] Step S1: Calculate the binary classification accuracy (Accuracy, Acc) between the pseudo-sequence and the original sequence under each SL, and calculate the embedding rate of the pseudo-sequence under each SL. The embedding rate is expressed as Bits Per Nucleotide (BPN). The formula for calculating the Bits Per Nucleotide (BPN) is as follows:

[0100]

[0101] Among them, L MSG With L B These represent the number of bits of the secret information in binary and the total number of bases in the pseudo-sequence of the steganography carrier, respectively.

[0102] Step S2: Construct the Effective Bits Per Nucleotide (EBPN) for the decision function. The formula for calculating the EBPN value is as follows:

[0103] EBPN = 2*(1-Acc)*BPN.

[0104] The effective single-base bit capacity EBPN is used to determine the impact of the security and capacity of the steganalytic vector pseudo-sequence.

[0105] Step S3: Calculate the EBPN value of the pseudo-sequences obtained under different SL segmentation processes. The pseudo-sequence with the largest EBPN value is the best stegtext carrier.

[0106] Furthermore, in step S1: the formula for calculating the binary classification accuracy value Acc is:

[0107]

[0108] The steganalytic vector pseudo-sequence and the original sequence are designated as positive and negative samples, respectively. TP and TN represent the number of correctly judged samples in the positive and negative samples, respectively, while FP and FN represent the number of incorrectly judged samples C in the positive and negative samples, respectively.

[0109] The method for evaluating pseudo-sequences in the sequence-generating DNA steganography method described in this invention is performed according to the following steps:

[0110] Step 1: Selection of Steganography Evaluation Parameters: The stealth of steganography can be reflected in perceptual stealth and statistical stealth. DNA is a physical molecular-level carrier, and its perceptual stealth can be reflected by its physical parameters. Among them, the parameter GC content is the proportion of the total number of G and C bases to all bases in the DNA strand, and the parameter melting temperature is the temperature corresponding to half of the DNA double helix unwinding, which is often used for DNA alignment. This patent calculates the difference between these two physical parameters in the original DNA sequence and the stegographic vector pseudo-sequence: GC content deviation (GC... d ) and melting temperature deviation (Tm) d ), to reflect the perceptual concealment of pseudo-sequences. As for the statistical concealment of DNA chains, MBBeck in paper

[14] clearly pointed out that the abundance of dinucleotides can be used to reflect the statistical characteristics of DNA chains. That is, by analyzing whether the abundance of dinucleotides in DNA chains is abnormal, it can be analyzed whether the sequence contains secret information. However, the paper does not provide a clear evaluation scheme. Therefore, this patent proposes to add a parameter that can be clearly calculated: the KL divergence of the relative abundance of dinucleotides (KL). ρ *) is used to evaluate its statistical concealment.

[0111] Step Two: This invention proposes two new parameters: GC content deviation (GC... d ) and melting temperature deviation (Tm) d To evaluate the perceptual concealment of pseudo-sequences:

[0112] Among them, GC content deviation (GC) d The formula for calculating ) is:

[0113]

[0114] Among them, GC r GC f These represent the GC content in the original DNA sequence and the pseudo-sequence of the stegtext vector, respectively.

[0115] Melting temperature deviation (Tm) d The formula for calculating ) is:

[0116]

[0117] Among them, Tm r Tmf These represent the melting temperatures of the original DNA sequence and the stegtext vector pseudo-sequence, respectively.

[0118] Furthermore, in step two, a new parameter is proposed: the KL divergence of the relative abundance of dinucleotides (KL... ρ *) is used to reflect the differences in the distribution of dinucleotides in the vector, and its calculation formula is:

[0119]

[0120] in, and These represent the relative abundance of different dinucleotides X and Y among the sixteen dinucleotides in the original DNA sequence and the stegtext vector pseudo-sequence, respectively.

[0121] The concealment of steganography is manifested in two aspects: perceptual concealment and statistical concealment. Since DNA itself is a molecular-level carrier, "perception" is generally achieved through experimental measurement. Therefore, two physical quantities reflecting the physical properties of DNA—GC content and melting temperature—are commonly used for DNA alignment and DNA synthesis. Here, we calculated the differences in pseudo-sequences of these two indicators between the original and stegographic vectors to demonstrate its perceptual concealment. Regarding statistical concealment, since only one researcher (MBBeck) has explicitly pointed out that steganalysis can be performed by analyzing anomalies in dinucleotide content, but without specifying a concrete evaluation method, we proposed the KL divergence of dinucleotide relative abundance. Therefore, all three quantities reflect concealment.

[0122] The GC content of DNA refers to the proportion of total G and C bases to all bases. The melting temperature of DNA refers to the temperature at which the DNA double helix unwinds halfway. GC content deviation (GC...) d ) and melting temperature deviation (Tm) d The KL divergence of relative dinucleotide abundance is used to measure the differences between different molecular-level carriers, thereby reflecting the stealth of the steganalysis carrier. ρ *Measures the overall differences in the relative abundance of 16 dinucleotides. Example:

[0123] In this invention, two natural non-coding region DNA sequences (Accession: GCA_009497935, GCA_004006475) were randomly selected from the European Bioinformatics Institute (EBI) database as examples.

[0124] The two natural non-coding region DNA sequences selected in this invention, after preprocessing, yielded datasets of 4304 sequences (200 bases, GCA_009497935), 4348 sequences (198 bases, GCA_009497935), 4582 sequences (200 bases, GCA_004006475), and 4629 sequences (198 bases, GCA_004006475), respectively. Based on the parity of the SL (Sequence Number Segment), they were combined into sequences of either 198 or 200 bases per short chain.

[0125] like Figure 3 The diagram shown is an overall flowchart of the present invention. After preprocessing, a base unit dictionary (V) is established. b The resulting base units are processed by word embedding. After word embedding, 30 high-dimensional vectors are stacked and used as input for a single pass of the RNN network.

[0126] The RNN network parameters are as described above. The state of the last hidden layer is input into the fully connected layer, and the parameters of the fully connected layer are as described above.

[0127] like Figure 4 As shown, after Softmax normalization, V is obtained. b The set of probabilities of all base units in (P) t ).

[0128] V b Each base unit in the array has its own probability value, with the color intensity representing the probability level; the darker the color, the higher the probability.

[0129] Whether to continue adjusting the RNN network parameters is determined by whether the defined loss function value converges. This invention uses the AdamW optimizer and backpropagation algorithm to adjust the RNN network parameters.

[0130] Once the loss function converges, stop model training and save the current RNN network parameter model (M). lstm ), and begin embedding secret information and generating pseudo-sequences containing secret information.

[0131] like Figure 5 The diagram illustrates the process of generating pseudo-sequences for steganography vectors. Specifically, it involves generating pseudo-sequences for P... t Adaptive dynamic grouping is performed to obtain multiple probability groups G. n Each group contains several base units. (By P) t The maximum probability p in max Determine the number B of the currently embedded secret bits. S Starting from the beginning of the secret bitstream, the number is B.S The secret bits are converted to the decimal number n. d In the next adaptive dynamic grouping operation, this number will be used as the group number for dynamic grouping among the resulting multiple groups, i.e., G[n] d ]. G[n d After probability normalization, adaptive dynamic grouping is performed. This process is repeated until the maximum normalized probability in the next group is greater than 0.5, at which point adaptive grouping stops. In the last group G... final The algorithm outputs a base unit based on probability sampling, and then sequentially links it to the generated pseudo-sequence S. p Afterwards, update S p The pseudocode for this process is as follows: Figure 1 The algorithm shown.

[0132] The process of extracting secret information is as follows: When an authorized recipient obtains the steganographic carrier pseudo-sequence S... p Then, first, S is divided according to the segmentation length of SL. p The process involves dividing the data to obtain the resulting carriers. Will Word embeddings are used and input into the RNN model M. lstm In this process, we obtain the base dictionary V. b The probability set P of all base units in t Applying adaptive dynamic coding to P t This yields multiple probability groups. Based on... The last base unit B last Group number n in the current group d Converting it to binary represents the partial secret bit hidden within this base unit. For B... last The grouping is processed by probability normalization before adaptive dynamic grouping. If, after grouping, B last If the maximum value of the group after probability normalization is still less than 0.5, repeat the above secret information extraction process; otherwise, stop extracting secret information and delete B. last ,renew The algorithmic pseudocode for this process is as follows: Figure 2 Algorithm 2 shown:

[0133] The embedding rate of DNA steganography is expressed as single-base bit capacity (BPN), which represents the average number of secret bits carried by a single base in the generated pseudo-sequence.

[0134] Experimental analysis and verification:

[0135] This invention selects four of the latest DNA steganography methods and obtains corresponding pseudo-sequences according to their technical implementation processes. Three evaluation parameters are used for these four pseudo-sequences and the pseudo-sequences generated by the steganography method described in this invention: the KL divergence of the relative abundance of dinucleotides (KL... ρ *), GC content deviation (GC d ) and melting temperature deviation (Tm) d The study compared the concealment and anti-steganography capabilities of the sequence using three indicators: effective single base bits (EBPN), classification accuracy (Acc), and sequence statistical distribution difference (measured by SWD).

[0136] First, the embedding rate BPN of the control method is calculated for comparison with the effective single-base bit EBPN:

[0137] According to the Repeated Bases Substitution (RBS) method proposed by C. Guo et al. in paper [3], this method searches for repeated base pairs in the original DNA sequence and performs base substitution at the second position of the repeated base pair. For example, in the base pair "AA" and "AAAA", the former can be substituted with the second A, and the latter can be substituted with the second and fourth A. The embedding rate of this method is determined by the natural DNA sequence being substituted. This invention uses this algorithm to replace the selected natural DNA and compares the base modification rate (Rate) of the pseudo-sequence after substitution with that of the natural DNA sequence. m Since bases can be viewed as quaternary codes, a base substitution can embed two bits. If all substituted bases are used to encode secret information, the embedding rate is twice the modification rate. The conversion formula is: BPN = Rate. m After conversion, the embedding rate of this method is 0.36.

[0138] According to the Table Lookup Substitution Method (TLSM) proposed by J.-S. Taur et al. in paper [2], this method generates pseudo-sequences by substituting bases at random positions in natural DNA after defining the binary numbers represented by the substitutions between bases (for example, "A->T" represents the binary number "10", and "G->C" represents the binary number "11"). This method can be equivalent to a random substitution, but its embedding rate is not fixed. The more bases substituted, the higher the embedding rate. Based on the implementation details of this method, the present invention adjusts its embedding rate to the same level as the present invention, which is 0.20.

[0139] According to the Generic Complementary Base Substitution (GCBS) method proposed by Amal Khalifa et al. in paper [5], this method encodes the secret information into a DNA sequence using binary encoding rules (A-00, C-01, G-10, T-11). The DNA sequence is then encrypted with a key and randomly divided into several segments. These segments are then randomly inserted into a natural DNA of equal length to obtain a pseudo-sequence. Since the length ratio of the natural DNA to the synthesized DNA sequence is 1:1, and the synthesized DNA sequence is randomly inserted into the natural DNA, this method is equivalent to Rate. m The random substitution rate is 50%. It's important to note that because this method introduces unencoded bases during key encryption, the embedding rate cannot be directly obtained from the conversion formula. Using the number of bits of secret information given in the article and the number of bases in the final generated steganalytic pseudo-sequence, the actual embedding rate BPN of this method is 0.4.

[0140] According to the Codon Content Restriction Synthesis (CCRS) proposed by JYChun et al. in paper [8], this method first converts the secret information into binary data and divides it into groups of five binary numbers. Based on the distribution of the total 64 codon contents in the natural DNA sequence, after removing the four start codons and stop codons, 32 codons are dynamically selected to encode these consecutive five-bit binary numbers. Finally, several of the remaining 28 start codons are randomly inserted into the encoded sequence so that the generated pseudo sequence has a similar codon content to the original DNA sequence. This invention reproduces the algorithm using the same steps, and the resulting embedding rate BPN is 0.89.

[0141] like Figure 6 The figure shows the names of four methods used as a comparison and their corresponding embedding rates (BPN).

[0142] Subsequently, based on the implementation details and technical specifications of the four methods described above, this invention reproduced their algorithms and conducted repeated experiments for each method, generating 10 different pseudo-sequences for each method. The steganalysis resistance and stealth capabilities were then compared with those of the 10 different pseudo-sequences generated using the steganography method described in this invention.

[0143] Then, the stealth and anti-steganography capabilities of the pseudocode generated by these five steganography methods are compared:

[0144] Steganography is used to identify whether a suspicious input sequence contains secret information. With the development of machine learning and deep learning, steganalysis techniques are constantly improving. A good steganalysis method must have strong resistance to the latest steganalysis techniques. Therefore, this embodiment selects three newly released steganalysis techniques as a steganalysis classification network to determine whether a pseudo-sequence has sufficient resistance to steganalysis.

[0145] The selected steganalysis techniques include: machine learning-based random forest steganalysis model, deep learning-based TEXTCNN steganalysis network, and deep learning-based RNNs steganalysis network proposed by Ho Bae et al. in the recently reported paper [9].

[0146] For the three steganalysis networks described above, the sample ratio between the original sequence and the generated pseudo-sequence is 1:1. The steganalysis resistance is represented by the classification accuracy Acc. The range of Acc is 50% to 100%, with the best Acc being 50%, indicating that the steganalysis network cannot distinguish the pseudo-sequence at all.

[0147] like Figure 7 The table shows the results of the steganalysis resistance between the pseudo-sequence generated by the control method and the pseudo-sequence generated by this invention. The black numbers in the table represent the average Acc as a percentage, and the lower right corner shows the standard deviation of Acc in repeated experiments. It can be seen that under the three steganalysis networks, the Acc of the control method is close to 100%, while the Acc of this invention ranges from 67.54% to 72.75%, significantly improving the steganalysis resistance. As an alternative implementation, this invention uses principal component analysis (PCA) to visualize the statistical distribution of the pseudo-sequence and the original DNA sequence through dimensionality reduction.

[0148] like Figure 8 The image shows the results of dimensionality reduction visualization of the pseudo-sequence and the original DNA sequence. The circle scatter plot represents the statistical distribution of the original DNA sequence, and the triangle scatter plot represents the statistical distribution of the pseudo-sequence. The higher the similarity between the two scatter plots, the closer the statistical distribution of the steganalytic sequence is to the original DNA sequence.

[0149] As an alternative implementation, this invention uses the Sliced-Wasserstein Distance (SWD) to quantify the statistical distribution difference between the original DNA sequence and the pseudo-sequence in the visualization results. The formula for calculating SWD is as follows:

[0150]

[0151] Where L represents the number of projection axes, P and Q represent the statistical distributions of the original DNA sequence and pseudo-sequence, respectively, and (θl)p and (θl)q represent the projections of P and Q onto the one-dimensional axis l. EMD stands for Earth Mover's Distance, which represents the minimum cost required to transform one distribution into another. However, directly calculating EMD is computationally unacceptable for large samples. Therefore, the sample is projected onto several axes, and the one-dimensional EMD is calculated separately, followed by the average, i.e., SWD. This method has a much lower computational cost than directly calculating EMD, and when the number of projection axes is not too small, the calculated result is very close to the EMD.

[0152] As an alternative implementation, the number of projection axes is 1000 when calculating SWD.

[0153] like Figure 9 The image shows the results of the analysis of the concealment of pseudo-sequences and original DNA sequences.

[0154] Among them, the ability to perceive concealment is an indicator (GC) proposed in this invention. d Tm d This is reflected in GC. d Tm d The smaller the value, the smaller the perceptible difference between the pseudo-sequence and the original DNA sequence, and the better the perceptible concealment of the pseudo-sequence. Statistical concealment is determined by the quantitative index SWD after dimensionality reduction visualization and the KL divergence (KL) of the relative abundance of dinucleotides. ρ *) to reflect SWD and KL ρ The smaller the value, the smaller the statistical difference between the pseudo sequence and the original DNA sequence, and the better the statistical concealment of the pseudo sequence.

[0155] like Figure 9 As shown, the black numbers represent the average of multiple pseudo-sequence indicators generated using the same method, and the number in the lower right corner represents the standard deviation. It can be seen that for different samples, the GC of this invention... d The average values ​​were 0.042% and 0.39%, Tm d The average values ​​were 0.011% and 0.099%, respectively, compared to the GC of the comparative methods. d Average range and Tm d The average values ​​range from 2.18% to 10.03% and from 0.38% to 2.17%, respectively. This invention significantly improves the perceptual concealment of pseudo-sequences. The average SWD values ​​of this invention are 33.40 and 77.5, respectively. ρ *(*10 -3 The average values ​​for KL were 0.24 and 0.12, while the average values ​​for SWD using the comparative method ranged from 54.18 to 113.58 and from 90.24 to 169.88.ρ *(*10 -3 The average values ​​range from 1.53 to 101.18 and from 1.09 to 102.00, and this invention significantly improves the statistical concealment of pseudo-sequences.

[0156] like Figure 10 The image shows a comparison of the hidden effective single base bits between the pseudo-sequence and the original DNA sequence. It can be seen that, for different samples, the effective single base bits (EBPN*10) of this invention... -2 The effective single base bits (EBPN*10) were 10.91 and 9.80, respectively, while the effective single base bits (EBPN*10) of the comparative method were... -2 The range is 0.079–2.93 and 0.010–3.97. It can be seen that this invention significantly improves the effective single-base bits of the pseudo-sequence.

[0157] In summary, the pseudo-sequences generated by the steganography method described in this invention have significant advantages over the pseudo-sequences generated by the four control methods in terms of concealment and resistance to steganography analysis.

[0158] References:

[0159] [1] Shiu, Hung-Jr, et al. "Data hiding methods based upon DNA sequences." Information Sciences 180.11(2010):2196-2208.

[0160] [2]Taur, Jin-Shiuh, et al. "Data hiding in DNA sequences based on tablelookup substitution." International Journal of Innovative Computing, Information and Control 8.10 (2012): 6585-6598.

[0161] [3]Guo, Cheng, Chin-Chen Chang, and Zhi-Hui Wang. "A new data hiding scheme based on DNA sequence." Int.J.Innov.Comput.Inf.Control 8.1(2012):139-149.

[0162] [4]Khalifa,Amal."LSBase:A key encapsulation scheme to improve hybridcrypto-systems using DNA steganography."2013 8th International Conference onComputer Engineering&Systems(ICCES).IEEE,2013.

[0163] [5]Khalifa,Amal,Ahmed Elhadad,and Safwat Hamad."Secure blind datahiding into pseudo DNA sequences using playfair ciphering and genericcomplementary substitution."Appl Math 10.4(2016):1483-1492.

[0164] [6]Clelland,Catherine Taylor,Viviana Risca,and Carter Bancroft."Hiding messages in DNA microdots."Nature 399.6736(1999):533-534.

[0165] [7]Smith,Geoff C.,et al."Some possible codes for encrypting data inDNA."Biotechnology letters 25.14(2003):1125-1130.

[0166] [8]Chun,Ji Young,Hye Lim Lee,and Ji Won Yoon."Passing go with DNAsequencing:Delivering messages in a covert transgenic channel."2015 IEEESecurity and Privacy Workshops.IEEE,2015.

[0167] [9]Bae,Ho,et al."DNA privacy:analyzing malicious DNA sequences usingdeep neural networks."IEEE / ACM Transactions on Computational Biology andBioinformatics(2020).

[0168]

[10] Bornholt,James,et al."A DNA-based archival storage system."Proceedings of the Twenty-First International Conference on ArchitecturalSupport for Programming Languages and Operating Systems.2016.

[0169]

[11] Srivastava,Nitish,et al."Dropout:a simple way to prevent neuralnetworks from overfitting."Thejournal ofmachine learning research 15.1(2014):1929-1958.

[0170]

[12] Cook,John D."Basic properties ofthe soft maximum."(2011).

[0171]

[13] Loshchilov,Ilya,and Frank Hutter."Fixing weight decayregularization in adam."(2018).

[0172]

[14] Beck,Marc B.,et al."Decoding methods for DNA steganalysis."6thInternational Conference on Bioinformatics and Computational Biology(BICoB).2014.

[0173]

[15] Chen,Yahui.Convolutional neural network for sentenceclassification.MS thesis.University of Waterloo,2015.

Claims

1. A sequence-generating DNA steganography method, characterized in that: The sequence-generating DNA steganography method based on conditional probability adaptive coding (ADG) utilizes an RNN network to learn the statistical distribution characteristics of the original DNA, and then uses an adaptive grouping algorithm to generate a pseudo-sequence of the steganalytic vector that is highly similar to the natural DNA and contains secret information. The pseudo-sequence is referred to as the pseudo-sequence. Finally, through biosynthesis technology, a DNA strand containing secret information is obtained by synthesizing according to the pseudo-sequence. This DNA strand has high perceptual concealment and statistical concealment, which significantly improves the resistance to steganalysis of the steganalytic DNA strand, making the information hiding of the steganalytic DNA strand highly secure. The sequence-generating DNA steganography method includes the following steps: Step 1: Preparation phase: Select the DNA sequence used for RNN network learning, which is the non-coding region of a natural DNA sequence; Step 2: Preprocessing stage: The selected DNA sequence is divided into segments of different lengths (SL) for preprocessing to obtain a dictionary of base units corresponding to the DNA sequence. The value of the segment length is the SL value. Step 3: Construct a dataset from the preprocessed DNA sequences; Step 4: Training phase of steganalytic vector pseudo-sequence model: Construct hidden layers and fully connected layers of RNN neural network, where RNN neural network is used to learn the statistical distribution characteristics of the dataset obtained in step 3; The hidden layers of an RNN (Recurrent Neural Network) preserve the learned statistical distribution by updating neuron weights and are used to generate pseudo-sequences; the fully connected layers of an RNN act as receivers of the RNN's output, receiving the output and ultimately outputting a dictionary of base units. The probability set is obtained by normalizing the fractions of all base units in the set. , used to calculate the loss function; Step 5: Construct the loss function. Whether training continues depends on whether the loss function converges: Adjust the neuron weight parameters of the RNN hidden layer; that is, adjust the base unit probability distribution obtained in Step 4. The parameters are input into the loss function, and the parameter values ​​of the neural network are adjusted based on whether the loss function converges. When the loss function converges, the trained steganalytic vector pseudo-sequence generation model is obtained, and the process proceeds to step 6. Otherwise, step 5 is repeated until the loss function converges. Step 6: Secret Information Embedding Stage: Construct a pseudo-sequence containing secret information based on the generative model obtained in Step 5. Specifically, this involves using the base unit dictionary obtained in Step 2. The probability set of all base units in the DNA is encoded using the ADG coding algorithm to generate a pseudo-sequence containing secret information. This pseudo-sequence containing secret information is then used to synthesize a steganographic DNA strand using biosynthesis technology, which can then be sent to the recipient.

2. The sequence-generating DNA steganography method according to claim 1, characterized in that: The decryption method, or the revealing step, is considered the reverse of the encryption process, and the key is the same, i.e., a symmetric encryption algorithm; the aforementioned key is the DNA steganalytic vector pseudo-sequence; the revealing method is as follows: Step C1: Steganographic DNA Strand Transmission Stage: The sender transmits the stegographic DNA strand to the receiver using any method suitable for transmitting natural DNA strands; simultaneously, to facilitate the receiver's retrieval of the secret information, the sender must also send the receiver a DNA strand synthesized from a pseudo-sequence containing the secret information, a training model and SL values, and a base unit dictionary. ; Step C2: Secret Information Extraction Stage: The process of information extraction and information hiding is reversed. Therefore, after receiving the steganalytic DNA strand, the receiver uses biotechnology to reverse-determine and analyze it to obtain the sequence of the DNA strand, i.e., the pseudo sequence containing the secret information. The pseudo sequence is then divided according to the SL value to obtain the processed sequence composed of base units. The base unit is an oligonucleotide with the number of bases SL. Step C3: Input the sequence obtained in step C2 into the received training model to obtain... Probability distribution of all base units ; Step C4: Use the ADG encoding algorithm to... Encode the information by extracting the secret information contained in the last word of the base unit sequence. After extraction, delete the last base unit, update the base unit sequence, and repeat steps C3 and C4 until all information in all base units has been extracted. At this point, the receiver has obtained all the secret information.

3. The sequence-generating DNA steganography method according to claim 1, characterized in that: In step 1: The DNA sequences used for learning the RNN network are all non-coding DNA sequences.

4. The sequence-generating DNA steganography method according to claim 1, characterized in that: In step 2: the specific method for preprocessing the DNA sequence is to divide the DNA sequence into different segment lengths, and the segment length is denoted as SL; During the partitioning process, the base sequence of the DNA sequence is preserved. After partitioning, each SL base is considered as a base unit, and a base unit dictionary corresponding to the DNA sequence is established. .

5. The sequence-generating DNA steganography method according to claim 1, characterized in that: In step 3, the dataset corresponding to the DNA sequence is obtained as follows: based on the arrangement order of the base units after division in step 2, the base units are combined into two or more short base chains; the aforementioned short base chains are of equal length, and the maximum length of a single short base chain is less than 200; thus, the dataset corresponding to the DNA sequence is obtained.

6. The sequence-generating DNA steganography method according to claim 1, characterized in that: In step 4, the RNN neural network uses a Long Short-Term Memory (LSTM) unit. This neural network contains two hidden layers, each with 512 LSTM units.

7. The sequence-generating DNA steganography method according to claim 1, characterized in that: In step 5, the loss function formula is: ; in, It is a tag for base units. It is the probability value corresponding to that base unit. It is the number of base units in a single short chain of bases.

8. A sequence-generating DNA steganography method according to any one of claims 1 to 7, characterized in that: The pseudo-sequence is obtained through the following steps: Step S1: Calculate the binary classification accuracy of the pseudo-sequence and the original sequence under each SL. Calculate the embedding rate of the pseudo-sequence under each SL, expressed as the number of base bits. Indicates single-base bit capacity. The calculation formula is: ； in, and These represent the number of bits of the secret information in binary and the total number of bases in the pseudo-sequence of the steganography carrier, respectively. Step S2: Constructing the effective single-base bit capacity The EBPN value is calculated using the following formula for the determination function: ； The effective single base bit capacity EBPN is used to determine the impact of the security and capacity of the stegtext vector pseudo-sequence. Step S3: Calculate the pseudo-sequences obtained under different SL segmentation processes. EBPN value, EBPN The one with the largest value is the best steganalysis pseudo-sequence.

9. A method for evaluating pseudo-sequences in a sequence-generating DNA steganography method according to any one of claims 1 to 7, characterized in that: Follow these steps: Step 1: Selection of Steganography Evaluation Parameters: The stealth of steganography can be reflected in perceptual stealth and statistical stealth. DNA is a physical molecular-level carrier, and its perceptual stealth can be reflected by its physical parameters. Among them, the parameter GC content is the proportion of the total number of G and C bases to all bases in the DNA strand, and the parameter melting temperature is the temperature corresponding to half of the DNA double helix unwinding. By calculating the differences between these two physical parameters in the original DNA sequence and the stegographic vector pseudo-sequence: GC content deviation... and melting temperature deviation This reflects the perceptual concealment of pseudo-sequences; while the statistical concealment of DNA strands is reflected by the KL divergence of the relative abundance of dinucleotides. To evaluate; Step 2: Add two new parameters: GC content deviation and melting temperature deviation To evaluate the perceptual concealment of pseudo-sequences: Among them, GC content deviation The calculation formula is: ； in, These represent the GC content in the original DNA sequence and the pseudo-sequence of the stegtext vector, respectively. Melting temperature deviation The calculation formula is: ； in, These represent the melting temperatures of the original DNA sequence and the stegtext vector pseudo-sequence, respectively.

Images