Asymmetric resource text steganography method and system based on double-scale language model

Abstract

The application discloses an asymmetric resource text steganography method and system based on a double-scale language model, and belongs to the technical field of information security. The method comprises the following steps: generating a steganography text subsequence token by token based on a double-scale large language model; and extracting secret information token by token based on a first language model. Under the premise that the steganography security meets the provable security requirement, the application can realize adaptive adjustment of embedding capacity, reduce the calculation complexity of the decoding end, and thus improve the practicability and scalability of the steganography communication system.

Description

Technical Field

[0001] This invention belongs to the field of information security technology, and in particular relates to an asymmetric resource text steganography method and system based on a dual-scale language model. Background Technology

[0002] Steganography is a technique that hides secret information within seemingly normal media to achieve covert communication. Unlike encryption, steganography not only focuses on the confidentiality of the message content but also emphasizes the imperceptibility of the communication itself. Even if an attacker performs statistical or semantic analysis on the communication content, it is difficult to determine whether secret information exists. Therefore, steganography has significant applications in covert communication, information warfare, and the transmission of sensitive information.

[0003] Traditional steganography methods are typically based on a carrier modification paradigm, embedding secret information by making minor modifications to multimedia carriers such as images, audio, or video. However, such methods inevitably compromise the statistical properties of the original carrier. With the development of steganalysis techniques, especially those based on machine learning and deep learning, traditional carrier modification-based steganography methods have gradually revealed insufficient security in the context of high-dimensional statistical features and end-to-end detection frameworks, severely limiting their covert communication capabilities.

[0004] To overcome the inherent security limitations of traditional carrier-modified steganography methods, Provably Secure Steganography (PSS) has been proposed. This type of method formally models steganographic security from a cryptographic perspective, reducing the security of the steganographic system to a problem of indistinguishability under standard cryptographic assumptions, theoretically guaranteeing the statistical consistency between the steganographic output and the natural channel output. Early constructs of provably secure steganography relied on the ideal random channel assumption, making them difficult to deploy in practical systems.

[0005] With the development of generative models, especially autoregressive (AR) language models, researchers have gradually combined provably secure steganography with natural language generation, proposing various text steganography methods based on language models. Compared to traditional carrier-modification steganography, language model-based steganography methods directly generate token sequences that conform to the distribution of natural language to carry secret information, avoiding modification of the existing carrier. This gives them a natural advantage in terms of semantic plausibility and statistical consistency, providing a new technical path for the engineering implementation of provably secure steganography.

[0006] Existing provably secure text steganography methods can be mainly categorized into two technical approaches. The first type of method is based on unbiased functions or approximately unbiased functions. The core idea of ​​the unbiased function construction is to divide or map the candidate symbols generated by the language model into multiple equally probable subsets, so that the embedded secret information is statistically uniformly distributed, thereby ensuring that the generated text is indistinguishable from the distribution of the original language model.

[0007] Under this technical approach, researchers have proposed several typical methods. For example, the ADG (Adaptive Dynamic Grouping) steganography method groups the distribution of candidate tokens, dividing the candidate symbol set into several groups with approximately equal probability and quality to carry secret information; the Meteor steganography method is based on arithmetic coding, which achieves an unbiased correspondence between secret bits and generated symbols by encoding and mapping the probability distribution of a language model; the Discop (DistributionCopies) steganography method embeds information by constructing multiple statistically equivalent distribution copies; and the SparSamp (SparseSampling) steganography method is based on sparse sampling techniques, which reduces sampling complexity and approximately achieves unbiased mapping by sparsifying the candidate symbol space. Although these methods differ in their technical implementation, they all essentially belong to the category of provably secure steganography construction based on unbiased functions.

[0008] Unbiased function-based steganography methods typically offer high embedding capacity and superior generation quality. However, their implementation relies on both communicating parties rigorously replicating the language model's sampling process during embedding and extraction phases, including candidate symbol sorting, probability normalization, and synchronized control of sampling randomness. This implies that the encoder and decoder need symmetrical computational resources and a consistent scale of language model parameters. In practical applications, in asymmetric resource environments—where the decoder runs on resource-constrained environments such as mobile terminals, embedded devices, or lightweight servers—the aforementioned symmetric resource assumption is often difficult to satisfy, severely limiting the practicality of such methods in asymmetric resource scenarios.

[0009] Another type of provably secure text steganography method is based on pseudo-random functions (PRFs) and rejection sampling mechanisms. By introducing key-related constraints during the generation process, the generated text strictly conforms to the distribution of the original language model while meeting embedding requirements. A significant advantage of this type of method is that the decoding end does not need to reproduce the language model generation process; information extraction can be completed simply by performing calculations on the generated symbols based on the shared key, making it naturally suitable for asymmetric resource scenarios. However, since the expected number of rejection samples increases exponentially with the number of embedded bits, this type of method presents a significant contradiction between embedding capacity and generation efficiency. As the embedding rate increases, the generation efficiency drops sharply, making it difficult to meet the requirements of high-capacity steganographic communication.

[0010] In summary, existing provably secure text steganography methods generally face the challenge of balancing embedding capacity, generation efficiency, and system deployability under asymmetric resource conditions. How to overcome the dependence on symmetric resources for unbiased function-based constructions while ensuring provably secure methods, and simultaneously overcome the inefficiency of rejection-sampling constructions at high embedding rates, is a key technical problem urgently needing to be solved in the field of text steganography. Summary of the Invention

[0011] The purpose of this invention is to provide an asymmetric resource text steganography method and system based on a dual-scale language model. By dividing the text steganography process into several clearly defined technical steps, while ensuring that the steganography security meets the provable security requirements, the method achieves adaptive adjustment of the embedding capacity and reduces the computational complexity of the decoding end, thereby improving the practicality and scalability of the steganography communication system.

[0012] To achieve the above objectives, the technical solution of the present invention includes the following:

[0013] An asymmetric resource text steganography method based on a dual-scale language model, applied to a steganography method, the method comprising: The first language model was jointly confirmed with the information extraction party; Based on the generated text symbols, obtain the text context of the current position. ; Based on the first language model, text context The analysis yields the steganographic capacity at the current location; Based on the steganography capacity, the secret bits to be embedded are sequentially extracted from the secret information stream; Based on a second language model, the secret bits to be embedded, and the text context. The text symbol at the current position is obtained; wherein the parameter size of the second language model is larger than that of the first language model; Re-execute the process of obtaining the text context of the current position based on the generated text symbols. The steganographic result of the secret information stream is obtained only after the secret information stream has been completely intercepted.

[0014] Furthermore, based on the first language model, the text context... The analysis yields the steganography capacity at the current location, including: Calling the first language model to the text context The analysis yields the first probability distribution of the candidate symbols; The uncertainty of the generated text at the current position is calculated based on the first probability distribution; The steganography capacity is determined based on the uncertainty of the generated text.

[0015] Furthermore, the uncertainty of the generated text at the current position is calculated based on the first probability distribution, including: Determine an indicator for measuring uncertainty, including an entropy function, a probability distribution concentration, or the number of valid candidates; Based on the first probability distribution and this indicator, the uncertainty of the generated text at the current position is obtained.

[0016] Furthermore, based on the second language model, the secret bits to be embedded, and the text context... Get the text symbol at the current position, including: Calling the second language model to the text context The analysis yields the second probability distribution of the candidate symbols; Rejection sampling is performed from the second probability distribution of candidate symbols based on a pseudo-random function to find a candidate symbol whose pseudo-random fingerprint is equal to the secret bit to be embedded; wherein, the pseudo-random function is generated based on a shared key jointly confirmed by the steganothermal party and the information extractor. Use the candidate symbol as the text symbol for the current position.

[0017] An asymmetric resource text information extraction method based on a dual-scale language model is applied to an information extraction method, the method comprising: The first language model is confirmed with the steganography party, and the steganography result generated by the asymmetric resource text steganography method based on the dual-scale language model according to any one of claims 1 to 4 is obtained; Based on the text symbols not extracted from the steganography results, the text context of the current position is obtained. ; Based on the first language model, text context The analysis yields the steganographic capacity at the current location; Based on the steganography capacity, extract the secret bit subsequence corresponding to the text symbol at the current position; Re-execute the step of obtaining the text context of the current position based on the text symbols not extracted from the steganography results. The process continues until all text symbols in the steganography result has been extracted, at which point the steganography extraction result is obtained.

[0018] Furthermore, based on the steganalysis capacity, the secret bit subsequence corresponding to the text symbol at the current position is extracted, including: Construct a pseudo-random function; wherein, the pseudo-random function is generated based on a shared key jointly confirmed by the steganography party and the information extraction party; The pseudo-random function is used to calculate the pseudo-random fingerprint of the text symbol at the current position. The pseudo-random fingerprint is truncated based on the steganography capacity to obtain the secret bit subsequence corresponding to the text symbol at the current position.

[0019] An asymmetric resource text steganography system based on a dual-scale language model, the method comprising: The first negotiation module is used to confirm the first language model with the information extractor. The first context generation module is used to obtain the text context of the current position based on the generated text symbols. ; The first capacity verification module is used to determine the text context based on the first language model. The analysis yields the steganographic capacity at the current location; The information interception module is used to sequentially extract the secret bits to be embedded from the secret information stream according to the steganography capacity; Secret steganography module for use with second language models, secret bits to be embedded, and text context. The text symbol at the current position is obtained; wherein the parameter size of the second language model is larger than that of the first language model; The first result generation module is used to obtain the steganographic result of the secret information stream after it has been completely intercepted.

[0020] An asymmetric resource text steganography extraction system based on a dual-scale language model, characterized in that it is applied to an information extraction method, the method comprising: The second negotiation module is used to jointly confirm the first language model with the steganography party and obtain the steganography result generated by the asymmetric resource text steganography method based on the dual-scale language model described above. The second context generation module is used to obtain the text context of the current position based on the text symbols that were not extracted in the steganography results. ; The second capacity verification module is used to determine the text context based on the first language model. The analysis yields the steganographic capacity at the current location; The secret extraction module is used to extract the secret bit subsequence corresponding to the text symbol at the current position based on the steganography capacity. The second result generation module is used to obtain the steganography extraction result after all text symbols in the steganography result have been extracted.

[0021] A computer device, comprising: a processor and a memory storing computer program instructions; wherein the processor, when executing the computer program instructions, implements the asymmetric resource text steganography method based on a dual-scale language model or the asymmetric resource text information extraction method based on a dual-scale language model as described above.

[0022] A computer-readable storage medium, characterized in that the computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the asymmetric resource text steganography method based on a dual-scale language model or the asymmetric resource text information extraction method based on a dual-scale language model as described above.

[0023] Compared with the prior art, the present invention has at least the following beneficial effects.

[0024] (1) Improve the feasibility of steganography in asymmetric resource scenarios. This invention introduces a dual-scale language model structure, which enables the decoding end to complete the extraction of steganographic information by running only a lightweight language model, which significantly reduces the computation and storage resource requirements of the decoding end and is suitable for resource-constrained practical application scenarios; (2) Improve text steganography embedding capacity. This invention dynamically allocates the number of embedding bits based on the candidate symbol entropy feature, so that the embedding capacity can be adaptively adjusted with the text generation context, which significantly improves the steganography capacity per unit symbol while ensuring security.

[0025] (3) Guaranteeing provable security. The present invention adopts a steganography construction method based on pseudo-random functions and rejection sampling. Under the standard cryptographic assumptions, it can be proven that the steganography security requirements are met, making the generated steganographic text statistically indistinguishable from the distribution of natural text.

[0026] (4) It has good model versatility and scalability. This invention does not depend on a specific language model structure or parameter size, and can be adapted to generative language models of different sizes and types, and has good engineering practicality and scalability potential. Attached Figure Description

[0027] Figure 1This is a flowchart of the asymmetric resource text steganography method based on a dual-scale language model of the present invention.

[0028] Figure 2 This is a flowchart of the encoding end steganography generation method of the present invention.

[0029] Figure 3 This is a flowchart of the method for extracting secret information from the receiving terminal according to the present invention. Detailed Implementation

[0030] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, so that those skilled in the art can have a more thorough understanding of the features and advantages of the present invention.

[0031] like Figure 1 As shown, the present invention comprises two parts: generating steganographic text subsequences per token based on a dual-scale large language model, and extracting secret information per token based on a first language model.

[0032] Step 1: Generate steganographic text subsequences token by token based on a dual-scale large language model. .

[0033] The steganography generation process of this invention adopts a token-by-token cyclic generation method, such as... Figure 2 As shown, it includes the following steps.

[0034] Step 1.1: Confirm the first language model together with the information extractor.

[0035] The first language model is an objective benchmark for bilateral synchronous embedding / extraction rules. It is used to determine rules based on context. Calculate uncertainty, and then use a function Determine the number of bits hidden at this location. Therefore, the steganography provider and the information extractor need to jointly confirm this beforehand.

[0036] Step 1.2: Obtain the text context of the current position based on the generated text symbols. .

[0037] In this embodiment, the text steganography process is performed by generating text symbols one by one. Assume the text consists of a discrete sequence of symbols, and in the... At each generation location, the currently generated text context can be represented as: in Indicates the first A generated text symbol.

[0038] Step 1.3: Analyze the text context based on the first language model The analysis yields the steganographic capacity at the current location.

[0039] This invention calls the first language model For the current text context To conduct analysis. Specifically, within a given context. Under the given conditions, the first probability distribution of the candidate symbols output by the first language model can be expressed as follows: in, Indicates candidate text symbols, This represents the first language model.

[0040] Based on the aforementioned candidate symbol probability distribution, the capacity estimation module calculates a metric reflecting the uncertainty of text generation. In this embodiment, the uncertainty metric can be characterized using an entropy function, for example, as follows:

[0041] It should be understood that the above uncertainty measurement method is only one example. In other implementations, uncertainty in text generation can also be characterized by probability distribution concentration, number of effective candidates or other equivalent indicators.

[0042] After obtaining the uncertainty measure value Subsequently, according to a preset capacity mapping rule, the present invention maps the uncertainty metric to the number of secret information bits that can be embedded at the current generation position. This mapping relationship can be abstractly represented as follows: in Indicates the location of generation Embeddable secret information bit length, function It is a monotonic mapping function. In this invention, it can be taken as... Implement a mapping from text generation uncertainty to steganalysis capacity.

[0043] Through the aforementioned capacity mapping mechanism, the embedding capacity can be dynamically changed with the uncertainty of text generation, thereby embedding more secret information in positions with higher uncertainty and reducing the embedding load in positions with lower uncertainty, so as to ensure the naturalness and stability of the generated text.

[0044] Step 1.4: Sequentially extract the secret bits to be embedded from the secret information stream according to the steganography capacity.

[0045] This invention operates from the beginning of the secret information stream. Before generating each text symbol, the steganalytic capacity at the current position has been determined using a first language model and a capacity mapping function. Then, the system sequentially extracts data of length from the current pointer of the information stream. A continuous bit segment is selected as the secret bit subsequence to be embedded in this round. After truncation, the pointer immediately moves forward. The bit points to the starting point of the next segment to be captured.

[0046] This process repeats itself. Due to the capacity of each location... The number of bits is dynamically changing (potentially 1, 2, 3... bits), so the length of the extracted subsequence also changes accordingly. The entire process ensures that the secret bits are extracted in strict order, without repetition or omission, preparing for subsequent steps to hide these bit segments one by one into the corresponding text symbols, until the entire information stream has been completely used.

[0047] Step 1.5: Based on the second language model, the secret bits to be embedded, and the text context This retrieves the text symbol at the current position.

[0048] In this embodiment, the steganography generation module calls the second language model. Generate the token actually used for the output. Second Language Model It is a generative language model with a large parameter scale, used to ensure that the generated text meets the needs of practical applications in terms of semantic coherence and language quality.

[0049] In a given text context Under these conditions, the second language model The second probability distribution for generating candidate symbols can be expressed as:

[0050] In this embodiment, the two communicating parties share a single key. And construct a pseudo-random function based on the key. This pseudo-random function maps candidate text symbols to bit strings of a specified length. At the generation location... The steganography generation module selects a length of [length missing] from the secret information stream to be embedded. The bit subsequence is used to map candidate text symbols using the pseudo-random function. When a candidate text symbol satisfies... When the condition is met, that is, the candidate symbol is considered to match the current secret message bit subsequence to be embedded, and the candidate symbol is... As the first Text symbols .

[0051] Wherein, the pseudo-random function A message authentication code-based approach is used to construct the system, meeting both security and efficiency requirements. Specifically, the communicating parties share a secret key. Under the premise of [specific conditions], the pseudo-random function can be constructed using a hash-based message authentication code (HMAC) algorithm. For any candidate text symbol... The output of the pseudo-random function can be expressed as: in, Indicated by key A hash message authentication code function with parameters.

[0052] In practical applications, candidate text symbols The data can be represented as a byte sequence or a bit sequence using a predetermined encoding method and used as the input message for the HMAC algorithm; the output of HMAC is a fixed-length bit string. Furthermore, the embedding capacity at the current generation position can be considered. Extract the first bit from the output bit string. The bit is used as a pseudo-random mapping result for matching the current generated position.

[0053] It should be understood that the above-described HMAC-based pseudo-random function construction method is only an exemplary implementation. Without deviating from the overall technical concept of this invention, other function construction methods that meet the pseudo-randomness requirements can also be used to achieve the same or similar technical effects.

[0054] Step 1.6: Re-execute the step of obtaining the text context of the current position based on the generated text symbols. The steganographic result of the secret information stream is obtained only after the secret information stream has been completely intercepted.

[0055] By repeating steps 1.2 to 1.5, the sequentially generated steganographic text symbols are concatenated to obtain a complete length of [length missing]. Steganographic text sequence .

[0056] Step 2: Based on the first language model The secret information is extracted one token at a time.

[0057] The steganography extraction process of this invention adopts a token-by-token loop approach, such as... Figure 3 As shown, it includes the following steps.

[0058] Step 2.1: Obtain the text context of the current position based on the unextracted text symbols in the steganography results. .

[0059] In this embodiment, the secret extraction process is performed using a text-by-text symbol generation method. Assume the text consists of a discrete sequence of symbols, and in the... At each generation location, the currently generated text context can be represented as: in Indicates the first A generated text symbol.

[0060] Step 2.2: Based on text context conditions Estimate the steganography capacity during steganography generation .

[0061] This invention calls the first language model In a given text context Under these conditions, the candidate distribution of the next text symbol is obtained to reproduce the capacity estimation results used by the encoder under the same context conditions. Specifically, the first language model in context... The first conditional probability distribution of the output candidate text symbols under given conditions can be expressed as:

[0062] Based on the candidate symbol distribution, the information extraction module calculates the uncertainty metric value for the corresponding generation position. In this embodiment, the uncertainty metric method is consistent with that of the encoding end; for example, it can be expressed in the form of information entropy.

[0063] In obtaining uncertainty metrics Subsequently, the information extraction module, using the same capacity mapping rule as the encoding end, maps the uncertainty measure to the embedding capacity corresponding to the current generation position, thereby obtaining the extractable secret bit length. The one used The process needs to be the same as the embedding process. Using the above method, the decoding end can reproduce the encoding end's output at the generation position based solely on the received text and the first language model, without requiring additional auxiliary information. The embedded capacity bit length used .

[0064] Step 2.3: Based on pseudo-random functions Perform secret message bit subsequence Extraction.

[0065] This invention determines the embedding capacity Then, using the key shared by both communicating parties. Call the pseudo-random function For the currently received text symbols Mapping is performed. In this embodiment, the pseudo-random function can be constructed using an HMAC-based approach, and its output is a fixed-length bit string. The information extraction module extracts the first... The bit, as the secret bit subsequence corresponding to the current text symbol, can be represented in the form of: It should be understood that since the encoding end uses the same context, capacity mapping rules, and shared key to perform pseudo-random function constraints during the steganography generation process, the decoding end can accurately recover the secret bit subsequence embedded by the encoding end at the corresponding generation position through the above mapping operation.

[0066] Each secret message bit subsequence The extraction steps include steps 2.1 to 2.3. By repeating the following steps, the secret bit subsequences corresponding to each generation position are concatenated according to the generation order, thereby recovering the complete length of... Secret information sequence .

[0067] In summary, this invention achieves an effective balance between theoretical security, embedding efficiency, text quality, and system practicality by organically combining rejection sampling construction, text generation uncertainty measurement, and a dual-scale language model collaborative mechanism, providing an engineering-feasible technical solution for provably secure text steganography. This invention introduces a rejection sampling mechanism based on the language model distribution during text generation, ensuring that the steganographic text strictly conforms to the output distribution of the original language model in a statistical sense. Specifically, the embedding of secret information is not achieved by directly modifying the generation probability or forcibly replacing text symbols, but rather by imposing constraints on candidate text symbols using a pseudo-random function constructed with a shared key, accepting the corresponding symbol as the generation result only when the constraints are met. Since this process essentially maintains the invariance of the language model distribution, the generated steganographic text is statistically indistinguishable from ordinary natural language text, thus theoretically meeting the basic requirements for provably secure steganography.

[0068] This invention employs a collaborative design of dual-scale language models, significantly reducing the computational complexity of the decoding end while ensuring text generation quality. Specifically, a second language model with a larger parameter scale is deployed only at the encoding end to generate high-quality, semantically coherent natural language text; while the decoding end only needs to rely on the first language model with a smaller parameter scale to complete the reproduction of embedding capacity and extraction of secret information. Through this asymmetric resource allocation method, this invention avoids the decoding end's dependence on a large-scale language model, reducing the overall computational and storage costs of the system, making it more suitable for resource-constrained receiving devices in practical communication scenarios. Furthermore, since the encoding and decoding ends use completely consistent rules in determining embedding capacity and mapping pseudo-random functions, this invention can reliably extract secret information without transmitting additional auxiliary information, ensuring the synchronization and robustness of the communication process. Even under long text generation or complex context conditions, the symbol-by-symbol cyclic embedding and extraction mechanism can maintain a stable working state, thus adapting to various practical application scenarios.

[0069] The above examples are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention. The scope of protection of the present invention shall be determined by the claims.

Claims

1. An asymmetric resource text steganography method based on a dual-scale language model, characterized in that, Applied to a steganography method, the method includes: The first language model was jointly confirmed with the information extraction party; Based on the generated text symbols, obtain the text context of the current position. ; Based on the first language model, text context The analysis yields the steganographic capacity at the current location; Based on the steganography capacity, the secret bits to be embedded are sequentially extracted from the secret information stream; Based on a second language model, the secret bits to be embedded, and the text context. The text symbol at the current position is obtained; wherein the parameter size of the second language model is larger than that of the first language model; Re-execute the process of obtaining the text context of the current position based on the generated text symbols. The steganographic result of the secret information stream is obtained only after the secret information stream has been completely intercepted.

2. The asymmetric resource text steganography method based on a dual-scale language model according to claim 1, characterized in that, Based on the first language model, text context The analysis yields the steganography capacity at the current location, including: Calling the first language model to the text context The analysis yields the first probability distribution of the candidate symbols; The uncertainty of the generated text at the current position is calculated based on the first probability distribution; The steganography capacity is determined based on the uncertainty of the generated text.

3. The asymmetric resource text steganography method based on a dual-scale language model according to claim 2, characterized in that, The uncertainty of the generated text at the current position is calculated based on the first probability distribution, including: Determine an indicator for measuring uncertainty, including an entropy function, a probability distribution concentration, or the number of valid candidates; Based on the first probability distribution and this indicator, the uncertainty of the generated text at the current position is obtained.

4. The asymmetric resource text steganography method based on a dual-scale language model according to claim 1, characterized in that, Based on a second language model, the secret bits to be embedded, and the text context. Get the text symbol at the current position, including: Calling the second language model to the text context The analysis yields the second probability distribution of the candidate symbols; Rejection sampling is performed from the second probability distribution of candidate symbols based on a pseudo-random function to find a candidate symbol whose pseudo-random fingerprint is equal to the secret bit to be embedded; wherein, the pseudo-random function is generated based on a shared key jointly confirmed by the steganothermal party and the information extractor. Use the candidate symbol as the text symbol for the current position.

5. A method for extracting asymmetric resource text information based on a dual-scale language model, characterized in that, Applied to an information extraction method, the method includes: The first language model is confirmed with the steganography party, and the steganography result generated by the asymmetric resource text steganography method based on the dual-scale language model according to any one of claims 1 to 4 is obtained; Based on the text symbols not extracted from the steganography results, the text context of the current position is obtained. ; Based on the first language model, text context The analysis yields the steganographic capacity at the current location; Based on the steganography capacity, extract the secret bit subsequence corresponding to the text symbol at the current position; Re-execute the step of obtaining the text context of the current position based on the text symbols not extracted from the steganography results. The process continues until all text symbols in the steganography result has been extracted, at which point the steganography extraction result is obtained.

6. The asymmetric resource text information extraction method based on a dual-scale language model according to claim 5, characterized in that, Based on the steganalysis capacity, extract the secret bit subsequence corresponding to the text symbol at the current position, including: Construct a pseudo-random function; wherein, the pseudo-random function is generated based on a shared key jointly confirmed by the steganography party and the information extraction party; The pseudo-random function is used to calculate the pseudo-random fingerprint of the text symbol at the current position. The pseudo-random fingerprint is truncated based on the steganography capacity to obtain the secret bit subsequence corresponding to the text symbol at the current position.

7. An asymmetric resource text steganography system based on a dual-scale language model, characterized in that, The method includes: The first negotiation module is used to confirm the first language model with the information extractor. The first context generation module is used to obtain the text context of the current position based on the generated text symbols. ; The first capacity verification module is used to determine the text context based on the first language model. The analysis yields the steganographic capacity at the current location; The information interception module is used to sequentially extract the secret bits to be embedded from the secret information stream according to the steganography capacity; Secret steganography module for use with second language models, secret bits to be embedded, and text context. The text symbol at the current position is obtained; wherein the parameter size of the second language model is larger than that of the first language model; The first result generation module is used to obtain the steganographic result of the secret information stream after it has been completely intercepted.

8. An asymmetric resource text steganography extraction system based on a dual-scale language model, characterized in that, Applied to an information extraction method, the method includes: The second negotiation module is used to jointly confirm the first language model with the steganography party and obtain the steganography result generated by the asymmetric resource text steganography method based on the dual-scale language model according to any one of claims 1 to 4. The second context generation module is used to obtain the text context of the current position based on the text symbols that were not extracted from the steganography results. ; The second capacity verification module is used to determine the text context based on the first language model. The analysis yields the steganographic capacity at the current location; The secret extraction module is used to extract the secret bit subsequence corresponding to the text symbol at the current position based on the steganography capacity. The second result generation module is used to obtain the steganography extraction result after all text symbols in the steganography result have been extracted.

9. A computer device, characterized in that, The computer device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the asymmetric resource text steganography method based on a dual-scale language model as described in any one of claims 1-4 or the asymmetric resource text information extraction method based on a dual-scale language model as described in any one of claims 5-6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the asymmetric resource text steganography method based on a dual-scale language model as described in any one of claims 1-4, or the asymmetric resource text information extraction method based on a dual-scale language model as described in any one of claims 5-6.

Images