Unmanned aerial vehicle secure communication text steganography and transmission method based on reinforcement learning

Abstract

This invention discloses a reinforcement learning-based method for secure text steganography and transmission in UAV communication, enabling secure transmission of control commands. Addressing the problems of easily eavesdropped and cracked control commands, weak anti-attack capabilities, and unstable communication links in existing UAV communication methods, this invention proposes a method integrating reinforcement learning and dynamic steganography generation. Through an adaptive dynamic steganography algorithm, control commands are embedded into the text. Reinforcement learning is used to optimize sentence length generation and corpus selection, ensuring the generated text's scene adaptability and resistance to steganography analysis. Simultaneously, transmission strategies are adjusted based on real-time channel conditions. Efficient parameter fine-tuning maintains the model's inference speed and low resource consumption, meeting the stringent requirements of real-time performance and stealth in UAV communication.

Description

Technical Field

[0001] This invention relates to the field of network security technology, and in particular to a method for secure text steganography and transmission in unmanned aerial vehicles based on reinforcement learning. Background Technology

[0002] With the rapid development of the low-altitude economy, drone swarms are increasingly widely used in civilian fields such as power line inspection and logistics delivery. Their core relies on collaborative communication between drones to achieve task division, attitude adjustment, and data sharing. However, current communication between drone swarms faces multiple threats, such as control commands being easily eavesdropped on and cracked, man-in-the-middle attacks, malicious attackers posing as legitimate nodes to access the swarm, and communication links being interfered with. Existing protection technologies are insufficient to fully protect the transmission of text-based control commands. For example, Chinese patent application publication number CN117376902A proposes a TPM-based drone command encryption method, which improves key security and command integrity, but the encrypted ciphertext still has sensitive transmission characteristics. Chinese patent publication number CN117220818A uses a low-overhead drone command communication method, converting text commands into simplified codes for transmission, reducing communication overhead, but lacks resistance to steganography. Chinese patent publication number CN113891312A employs a transmission strategy of first obfuscating and then encrypting command strings to achieve lightweight command concealment. Therefore, to address the security challenges in transmitting control commands between drone swarms, there is an urgent need for a method that can both achieve covert command transmission and ensure integrity and resistance to attacks.

[0003] Text steganography, an important branch of information hiding, aims to conceal secret information within publicly available "carrier text" through subtle manipulation of natural language text, achieving covert communication. Specifically, it ensures that the secret information can be extracted by the authorized party while maintaining that the carrier text and normal text are not significantly different in terms of statistical features and semantic coherence, thereby circumventing third-party steganalysis. Generative text steganography uses generative models to directly generate carrier text that appears natural but actually contains hidden secret information, avoiding statistical anomalies caused by the mechanical modification of existing text in traditional methods.

[0004] Large language models, with their superior contextual understanding, semantic reasoning, and natural text generation capabilities, provide strong technical support for generative text steganography. For example, Chinese patent application publication number CN 120163135A proposes a text steganography method that incorporates a large language model, utilizing a pre-defined knowledge graph text generation model to improve the quality and naturalness of text generation. However, generative steganography based on large language models still faces challenges. First, there is the efficiency of model adaptation. General-purpose large models are not optimized for steganography tasks, and direct use can easily compromise the naturalness of the text due to overfitting of secret information. If full fine-tuning is used to adapt the model to the steganography task, the computational resources consumed are too large, making it difficult to meet the computational limitations and real-time communication requirements of low-altitude UAVs. Second, traditional large language model-driven steganography methods usually rely on static corpora and fixed generation strategies, and cannot dynamically adjust steganography parameters according to the real-time channel state or task requirements of the communication environment, making it difficult to balance steganography efficiency and concealment in highly dynamic scenarios.

[0005] Reinforcement learning, as a machine learning paradigm, trains decision-making agents to interact with the environment and optimizes policy selection based on the real-time state of the environment and feedback from actions. Combining reinforcement learning with generative text steganography allows for consideration of real-time channel states, taking into account the context and sentence length of the generated text, thus improving the stealth of the steganography. Summary of the Invention

[0006] This invention provides a reinforcement learning-based method for secure communication text steganography and transmission for unmanned aerial vehicles (UAVs). This addresses the problems of existing semantic-based text steganography methods in low-altitude UAV swarm transmission scenarios, such as the vulnerability of control commands to eavesdropping and cracking, weak anti-attack capabilities, and unstable communication links. This invention utilizes fine-tuning techniques to adjust the parameters of a large language model and leverages reinforcement learning to optimize sentence length and corpus type, thereby improving the scene adaptability of the generated text and the model inference speed. This achieves real-time and covert communication between low-altitude UAV swarms.

[0007] A first aspect of this invention provides a method for secure text steganography and transmission in unmanned aerial vehicles (UAVs) based on reinforcement learning, comprising the following steps: 1) Load the pre-trained large language model, freeze the original parameters of the large language model, and fine-tune the parameters of the large language model based on the UAV command corpus to obtain a fine-tuned large language model adapted to the UAV communication scenario. 2) Deploy the finely tuned large language model on the edge computing device of the drone; 3) During the k-th communication, the length of the binary instruction data to be transmitted, the system task acquired by the UAV, the transmission delay and steganography delay of the previous time slot, the channel gain of the current time slot and the packet loss rate of the previous time slot are combined into the state variables of the current communication environment. 4) Construct a reinforcement learning Q-network, initialize the network weights, learning rate and discount factor, and calculate the optimal strategy based on the state variables using the reinforcement learning Q-network, outputting the corpus configuration and sentence length of the generated text; 5) Select the corresponding corpus according to the output corpus configuration, and use the fine-tuned large language model to calculate the conditional probability distribution of each word in the vocabulary in the current text; 6) Based on the conditional probability distribution, the vocabulary is adaptively divided into multiple groups with approximately equal probability sums; 7) Divide the binary secret information to be embedded into multiple binary sub-blocks, and map each binary sub-block to a corresponding grouping index; 8) Based on the grouping index, select the word with the highest probability from the corresponding group and generate the carrier text sequence in sequence until the sentence length is reached; 9) Detoxify the carrier text sequence into natural text, which will be used as the final text carrier for covert transmission.

[0008] Optionally, in one embodiment of the present invention, the specific process of parameter fine-tuning in step 1) is as follows: The weight matrices of the attention layer and feedforward layer of the pre-trained large language model are decomposed into low-rank matrices, which are the product of the projection matrix and the restoration matrix. The low-rank matrix is ​​trained using a drone command corpus, and the weight matrix is ​​frozen. After training, the weight matrix is ​​merged with the trained projection matrix and restoration matrix to obtain a finely tuned large language model adapted to drone communication scenarios.

[0009] Optionally, in one embodiment of the present invention, in step 3), the specific process of constructing the state vector is as follows: during the k-th communication, the UAV obtains the system task sent by the ground station. and the instruction data that needs to be transmitted. The length is The transmission delay of the previous time slot is calculated based on the round-trip delay when sending data packets. The packet loss rate for the previous time slot is obtained based on the packet loss statistics reported in the feedback. The time interval between the start of calling the large language model and the end of text generation is denoted as the steganography delay of the previous time slot. Then, estimate the current channel gain based on the real-time channel status. Construct the state vector at the k-th communication. .

[0010] Optionally, in one embodiment of the present invention, in step 4), the optimal policy output by the Q network is defined as... ,in, Representative corpus configuration, The sentence length of the generated text, k represents the k-th communication, and the drone will transmit the state vector. The input is fed into a Q-network, and the maximum value function is obtained according to optimization theory. The Q network outputs the corpus configuration used in the current state and the sentence length of the generated text.

[0011] Optionally, in one embodiment of the present invention, in step 6), the vocabulary grouping method includes: calculating the number of grouping levels U based on the maximum word probability, and calculating the target probability threshold for each group based on the sum of the conditional probabilities of all words in the candidate word set R. Then group the words using the following formula. ,in, This represents the optimal word chosen. This represents the sum of conditional probabilities in the current group. This represents the conditional probability distribution corresponding to each word; by iteratively selecting words to add to groups, the cumulative conditional probability of each group approaches the target probability threshold until all words are assigned, achieving approximately equal sums of conditional probabilities within each group.

[0012] Optionally, in one embodiment of the present invention, in step 8), the process of selecting the word with the highest conditional probability is as follows: Based on the binary sub-block of the current secret information Corresponding grouping index Filter out the target group from this group. words probability distribution Select the word with the highest conditional probability from this group. It is added to the carrier text sequence and the text carrier used for steganographic transmission is updated.

[0013] The reinforcement learning-based method for secure UAV communication text steganography and transmission in this invention has the following advantages: 1. This invention analyzes multi-dimensional task conditions such as communication channel status and task requirements, and dynamically selects suitable steganography parameters and generation strategies, effectively improving the real-time performance and concealment of steganography tasks. 2. By selecting a corpus and employing an adaptive embedding strategy, this invention embeds secret information while ensuring the semantic coherence of the generated text, effectively improving the naturalness of the carrier text and reducing the risk of being identified by steganography analysis. 3. This invention updates only a few key parameters and combines them with a dynamic generation strategy to reduce computing resource consumption while maintaining LLM inference speed. It effectively adapts to scenarios where the computing power of low-altitude UAV clusters is limited, and achieves efficient deployment of steganography tasks.

[0014] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a reinforcement learning-based method for secure communication text steganography and transmission for unmanned aerial vehicles (UAVs) according to an embodiment of the present invention. Detailed Implementation

[0016] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0017] This invention proposes a reinforcement learning-based method for secure text steganography and transmission in UAV communication, enabling secure transmission of control commands. Existing UAV communication methods suffer from vulnerabilities such as easily eavesdropped and cracked control commands, weak anti-attack capabilities, and unstable communication links. Covert communication, as a novel mode for ensuring wireless security, has attracted significant attention. This invention, based on an adaptive dynamic steganography algorithm, embeds control commands into the text. Reinforcement learning is used to optimize sentence length generation and corpus selection, ensuring the generated text's scene adaptability and resistance to steganography analysis. Furthermore, efficient parameter fine-tuning maintains the model's inference speed and low resource consumption, meeting the stringent requirements of real-time performance and covertness in UAV communication.

[0018] Figure 1 This is a flowchart illustrating a reinforcement learning-based method for secure communication text steganography and transmission for unmanned aerial vehicles (UAVs) according to an embodiment of the present invention.

[0019] like Figure 1 As shown, the reinforcement learning-based method for secure UAV communication text steganography and transmission includes the following steps: 1) Load the pre-trained large language model, freeze the original parameters of the large language model, and fine-tune the parameters of the large language model based on the UAV command corpus to obtain a fine-tuned large language model adapted to the UAV communication scenario. 2) Deploy the finely tuned large language model on the edge computing device of the drone; 3) During the k-th communication, the length of the binary instruction data to be transmitted, the system task acquired by the UAV, the transmission delay and steganography delay of the previous time slot, the channel gain of the current time slot and the packet loss rate of the previous time slot are combined into the state variables of the current communication environment. 4) Construct a reinforcement learning Q-network, initialize the network weights, learning rate and discount factor, and calculate the optimal strategy based on the state variables using the reinforcement learning Q-network, outputting the corpus configuration and sentence length of the generated text; 5) Select the corresponding corpus according to the output corpus configuration, and use the fine-tuned large language model to calculate the conditional probability distribution of each word in the vocabulary in the current text; 6) Based on the conditional probability distribution, the vocabulary is adaptively divided into multiple groups with approximately equal probability sums; 7) Divide the binary secret information to be embedded into multiple binary sub-blocks, and map each binary sub-block to a corresponding grouping index; 8) Based on the grouping index, select the word with the highest probability from the corresponding group and generate the carrier text sequence in sequence until the sentence length is reached; 9) Detoxify the carrier text sequence into natural text, which will be used as the final text carrier for covert transmission.

[0020] Optionally, in one embodiment of the present invention, the specific process of parameter fine-tuning in step 1) is as follows: The weight matrices of the attention layer and feedforward layer of the pre-trained large language model are decomposed into low-rank matrices, which are the product of the projection matrix and the restoration matrix. The low-rank matrix is ​​trained using a drone command corpus, and the weight matrix is ​​frozen. After training, the weight matrix is ​​merged with the trained projection matrix and restoration matrix to obtain a finely tuned large language model adapted to drone communication scenarios.

[0021] Optionally, in one embodiment of the present invention, in step 3), the specific process of constructing the state vector is as follows: during the k-th communication, the UAV obtains the system task sent by the ground station. and the instruction data that needs to be transmitted. The length is The transmission delay of the previous time slot is calculated based on the round-trip delay when sending data packets. The packet loss rate for the previous time slot is obtained based on the packet loss statistics reported in the feedback. The time interval between the start of calling the large language model and the end of text generation is denoted as the steganography delay of the previous time slot. Then, estimate the current channel gain based on the real-time channel status. Construct the state vector at the k-th communication. .

[0022] Optionally, in one embodiment of the present invention, in step 4), the optimal policy output by the Q network is defined as... ,in, Representative corpus configuration, The sentence length of the generated text, k represents the k-th communication, and the drone will transmit the state vector. The input is fed into a Q-network, and the maximum value function is obtained according to optimization theory. The Q network outputs the corpus configuration used in the current state and the sentence length of the generated text.

[0023] Optionally, in one embodiment of the present invention, in step 6), the vocabulary grouping method includes: calculating the number of grouping levels U based on the maximum word probability, and calculating the target probability threshold for each group based on the sum of the conditional probabilities of all words in the candidate word set R. Then group the words using the following formula. ,in, This represents the optimal word chosen. This represents the sum of conditional probabilities in the current group. This represents the conditional probability distribution corresponding to each word; by iteratively selecting words to add to groups, the cumulative probability of each group approaches the target probability threshold until all words are assigned, achieving approximately equal probabilities within each group.

[0024] Optionally, in one embodiment of the present invention, the process of selecting the word with the highest probability in step 8) is as follows: Based on the binary sub-block of the current secret information Corresponding grouping index Filter out the target group from this group. words probability distribution Select the word with the highest conditional probability from this group. It is added to the carrier text sequence and the text carrier used for steganographic transmission is updated.

[0025] The following specific embodiment illustrates the reinforcement learning-based UAV secure communication text steganography and transmission method of the present invention, which includes the following steps:

[0026] Step 1: Load the pre-trained large language model and freeze the main parameters. The large language model used in this invention is Llama, so its computational efficiency matches that of the drone. Other lightweight large language models can also be used. Define the corpus configuration used for communication between drone clusters as follows: The total number of corpora is ,Right now For the attention layer and feedforward layer of a large language model, the original weights are... Introduce low-rank decomposition: ,in For the projection matrix, To restore the matrix, It is a low-rank dimension. Only [the following is included] , Set as trainable parameters. Select the corpus. Input model, use Adam optimizer to update only , Continue until the loss converges. Merge the original weights. With the low-rank matrix after training , This results in a finely tuned large language model adapted to the current communication scenario.

[0027] Step 2: Deploy the large language model with fine-tuned parameters on the drone edge computing device.

[0028] Step 3: During the drone's first... During this communication, the text carrier currently used for steganographic transmission will be denoted as... The binary command data that the drone needs to transmit is The drone receives system tasks sent from the ground station. and the instruction data that needs to be transmitted. The length is The transmission delay of the previous time slot is calculated based on the round-trip delay when sending data packets. The packet loss rate for the previous time slot is calculated based on the reported packet loss statistics and the number of data packets sent. Then, the time interval from the start of calling the large language model to the end of text generation is recorded as the steganography delay of the previous time slot. Estimate the current channel gain based on the real-time channel status. At this point, the above variables are constructed in the first... State vector during the second communication ,Right now .

[0029] Step 4: Initialize the vocabulary of the large language model Initial probability distribution It is uniformly distributed.

[0030] Step 5: Initialize the weight parameters of the reinforcement learning Q-network and initialize the learning rate. and discount factor .

[0031] Step 6: The UAV uses the acquired state vector The optimal strategy is calculated using a Q-network. , ,in , Representative corpus configuration, This represents the sentence length of the generated text.

[0032] Step 7: Use the serial number The corpus is used for forward inference of the large language model to calculate the vocabulary. Each word In the current text Conditional probability distribution ,in This represents the large language model's understanding of words given a context. Conditional probability distribution, generation probability arrangement ,satisfy .

[0033] Step 8: From the conditional probability distribution Choose the maximum probability Calculate the number of levels in the steganography group. , .

[0034] Step 9: Initialize the vocabulary grouping set Candidate word set .

[0035] Step 10: Based on the current candidate word set The sum of the conditional probabilities of all words is used to calculate the target probability threshold for each group in the i-th layer. , .

[0036] Step 11: If the current group The sum of probabilities is less than the target probability threshold. And the candidate word set If the value is not empty, continue selecting words to add to the group. The specific process is as follows: Select words from the candidate word set that sum the conditional probabilities of the current group. Conditional probability distribution of words The sum is closest to the target probability threshold words ,Right now Then add the word to the group. The word is then removed from the candidate word set to avoid duplicate selections.

[0037] Step 12: Repeat steps 10-11 until each group is finished. The cumulative probability sum is approximately equal to its corresponding target probability threshold. This enables adaptive probability-balanced partitioning of the vocabulary.

[0038] Step 13: Based on the number of group levels Calculate the length of each secret information block , will embed binary secret information Divided into multiple lengths binary sub-blocks Initialize the final generated carrier text sequence .

[0039] Step 14: Divide the binary sub-block Convert to grouped index Used to select pre-defined target groups .

[0040] Step 15: Define the probability distribution of constrained words for: ; That is, based on the conditional probabilities and indicator functions output by the large language model, groups belonging to the target group are selected. words probability distribution .

[0041] Step 16: Group from target Choose the word with the highest probability , Selected words Added to the carrier text sequence and the current text end.

[0042] Step 17: Repeat steps 14 to 16 until the length of the generated sequence meets the sentence length requirement of the generated text. .

[0043] Step 18: Sequence Converted into natural text through detoxification. Assign a value to As the final output carrier, it is used for covert transmission and is combined with real-time channel status. Select the appropriate transmission power to send the text to the receiving end.

[0044] The following detailed description of the UAV swarm secure communication method based on reinforcement learning and adaptive text steganography according to the present invention is illustrated through a specific embodiment.

[0045] Step 1: Tune the Llama model and deploy it to the computing devices of the drone swarm. Initialize the current text carrier. =“.

[0046] Step 2: Initialize the network weight parameters of the reinforcement learning Q-network and initialize the learning rate. Discount factor .

[0047] Step 3: In the During this communication, the UAV received the system mission sent by the ground station. and the instruction data that needs to be transmitted. The length is The transmission delay of the previous time slot is calculated based on the round-trip delay when sending data packets. The packet loss rate for the previous time slot is obtained based on the packet loss statistics reported in the feedback. The time interval between the start of calling the large language model and the end of text generation is denoted as the steganography delay of the previous time slot. Then, estimate the current channel gain based on the real-time channel status. At this point, the above variables are constructed in the first... State vector during the second communication ,Right now .

[0048] Step 4: Select the corpus configuration to use based on the output of the reinforcement learning Q-network. and the sentence length of the generated text The pre-trained corpora include Movies, SNS, and News, i.e. Therefore, before generating the probability distribution of each word using a large language model, it is necessary to select from three pre-trained large models based on corpora.

[0049] Step 5: Select the current text The last word Generate conditional probability distributions using large models Store all possible words in a candidate word set R, and select the one with the highest probability. Calculate the number of group levels .

[0050] Step 6: Based on the number of group levels Calculate the target probability threshold for each group Then, words are continuously selected from the candidate word set R to be added to the group. The process continues until the sum of the probabilities in the groups exceeds the target probability threshold. Or there are no words to choose from in the candidate word set R.

[0051] Step 7: Initialize the final generated carrier text sequence Calculate the length of each secret information block. and corresponding binary sub-blocks .

[0052] Step 8: Binary Sub-block Convert to grouped index From target group Choose the word with the highest probability Added to the carrier text sequence and the current text end.

[0053] Step 9: Repeat steps 5 through 8 until the carrier text sequence is complete. The sentence length required to generate the text is reached. Finally, a detoxing operation is performed to convert the sequence into text for transmission.

[0054] The UAV swarm secure communication method based on reinforcement learning and adaptive text steganography proposed in this invention considers both the computational efficiency of a large language model and the environmental adaptability of the generated text. It generates the optimal sentence length and corpus type for the current moment through dynamic decision-making by a reinforcement learning agent. Fine-tuning techniques are used to efficiently fine-tune the parameters of the large language model, ensuring the scene adaptability and naturalness of the generated text while maintaining the model's inference speed and low resource consumption.

[0055] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0056] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0057] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

Claims

1. A method for secure text steganography and transmission in unmanned aerial vehicles (UAVs) based on reinforcement learning, characterized in that, Includes the following steps: 1) Load the pre-trained large language model, freeze the original parameters of the large language model, and fine-tune the parameters of the large language model based on the UAV command corpus to obtain a fine-tuned large language model adapted to the UAV communication scenario. 2) Deploy the finely tuned large language model on the edge computing device of the drone; 3) During the k-th communication, the length of the binary instruction data to be transmitted, the system task acquired by the UAV, the transmission delay and steganography delay of the previous time slot, the channel gain of the current time slot and the packet loss rate of the previous time slot are combined into the state variables of the current communication environment. 4) Construct a reinforcement learning Q-network, initialize the network weights, learning rate and discount factor, and calculate the optimal strategy based on the state variables using the reinforcement learning Q-network, outputting the corpus configuration and sentence length of the generated text; 5) Select the corresponding corpus according to the output corpus configuration, and use the fine-tuned large language model to calculate the conditional probability distribution of each word in the vocabulary in the current text; 6) Based on the conditional probability distribution, the vocabulary is adaptively divided into multiple groups with approximately equal probability sums; 7) Divide the binary secret information to be embedded into multiple binary sub-blocks, and map each binary sub-block to a corresponding grouping index; 8) Based on the grouping index, select the word with the highest probability from the corresponding group and generate the carrier text sequence in sequence until the sentence length is reached; 9) Detoxify the carrier text sequence into natural text, which will be used as the final text carrier for covert transmission.

2. The method according to claim 1, characterized in that, In step 1), the specific process of parameter fine-tuning is as follows: The weight matrices of the attention layer and feedforward layer of the pre-trained large language model are decomposed into low-rank matrices, which are the product of the projection matrix and the restoration matrix. The low-rank matrix is ​​trained using a drone command corpus, and the weight matrix is ​​frozen. After training, the weight matrix is ​​merged with the trained projection matrix and restoration matrix to obtain a finely tuned large language model adapted to drone communication scenarios.

3. The method according to claim 1, characterized in that, In step 3), the specific process of constructing the state vector is as follows: During the k-th communication, the UAV receives the system task sent by the ground station. and the instruction data that needs to be transmitted. The length is The transmission delay of the previous time slot is calculated based on the round-trip delay when sending data packets. The packet loss rate for the previous time slot is obtained based on the packet loss statistics reported in the feedback. ; The time interval between the start of calling the large language model and the end of text generation is denoted as the steganography delay of the previous time slot. Then, estimate the current channel gain based on the real-time channel status. Construct the state vector at the k-th communication. .

4. The method according to claim 1, characterized in that, In step 4), the optimal policy output by the Q network is defined as... ,in, Representative corpus configuration, The sentence length of the generated text, k represents the k-th communication, and the drone will transmit the state vector. The input is fed into a Q-network, and the maximum value function is obtained according to optimization theory. The Q network outputs the corpus configuration used in the current state and the sentence length of the generated text.

5. The method according to claim 1, characterized in that, In step 6), the vocabulary grouping method includes: calculating the grouping level U based on the highest word probability, and calculating the target probability threshold for each group based on the sum of the conditional probabilities of all words in the candidate word set R. Then group the words using the following formula. ,in, This represents the optimal word chosen. This represents the sum of conditional probabilities in the current group. This represents the conditional probability distribution corresponding to each word; by iteratively selecting words to add to groups, the cumulative conditional probability of each group approaches the target probability threshold until all words are assigned, achieving approximately equal sums of conditional probabilities within each group.

6. The method according to claim 1, characterized in that, In step 8), the process of selecting the word with the highest conditional probability is as follows: Based on the binary sub-block of the current secret information Corresponding grouping index Filter out the target group from this group. words probability distribution Select the word with the highest conditional probability from this group. It is added to the carrier text sequence and the text carrier used for steganographic transmission is updated.

Images