A text steganalysis method, device, equipment, medium and program product

Abstract

This application provides a text steganography analysis method, apparatus, device, medium, and program product, which can be applied to the field of text detection. The method includes: in response to receiving text to be detected, inputting a first feature vector of the text to be detected into a semantic baseline extraction layer of an analysis model, extracting low-frequency semantic features and artifact dimension signal features from the first feature vector, and fusing the two to obtain a second feature vector; inputting the first feature vector into an artifact feature amplification layer to amplify the artifact dimension signal features in the first feature vector, obtaining a third feature vector including the amplified artifact dimension signal features; inputting the second feature vector and the third feature vector into a global representation extraction layer for feature fusion to obtain a fourth feature vector; inputting the fourth feature vector into a classification head module to obtain a detection result, the detection result including whether the text to be detected is normal text or steganographic text.

Description

Technical Field

[0001] This application relates to the field of text detection, specifically to a text steganography analysis method, apparatus, device, medium, and program product. Background Technology

[0002] Text steganography, a key technology for information security, is primarily used to detect whether hidden information is embedded in text. With the rapid development of generative text steganography, higher demands are placed on the accuracy, efficiency, and generalization ability of detection techniques. Currently, mainstream generative text steganography techniques are mainly divided into two categories: one is the AC and Discop algorithms based on probability distribution modification, which embed secret information by modifying the conditional probability distribution of words, easily causing slight shifts in the statistical distribution of the text compared to natural language; the other is the VS algorithm based on structural constraints, which encodes the secret information into global latent variables, easily leading to structural misalignments in the local context of the text. Both types of algorithms can generate semantically coherent text with statistical features close to natural language, significantly increasing the difficulty of detection.

[0003] Existing text steganalysis techniques are mainly divided into two categories: traditional deep learning methods and detection methods based on large language models, both of which have significant technical shortcomings. Traditional deep learning methods, such as TS-RNN and TS-FCN, rely solely on statistical features or shallow semantic inconsistencies for detection, lacking the ability to model deep, long-distance dependencies and subtle structural misalignments in the text. When faced with high-quality generative stegtext, they struggle to identify minute perturbations introduced by complex sampling strategies, resulting in low detection accuracy.

[0004] While detection methods based on large language models, such as GS-Llama, can achieve good detection results, they have several limitations: First, using a generative model with over 7 billion parameters as the foundation results in high computational costs and large memory consumption during training and inference, and the model's computational complexity increases quadratically with the length of the text sequence, making it difficult to deploy in resource-constrained environments. Second, using a decoder architecture designed for text generation to handle classification and detection tasks has inherent disadvantages in bidirectional contextual feature extraction and fine-grained discrimination, resulting in low detection efficiency. Third, using LoRA technology only as a general parameter adaptation tool without designing anti-overfitting strategies for steganography tasks makes the model prone to overfitting the semantic content of the training data rather than capturing steganography artifacts. When facing unknown steganography algorithms or cross-semantic domain detection, its performance drops significantly, and its generalization ability is insufficient.

[0005] Overall, existing text steganalysis techniques struggle to balance detection accuracy and computational efficiency, and their cross-domain and cross-algorithm generalization capabilities are poor, failing to meet the practical application requirements for high-precision, high-efficiency, and highly robust steganalysis detection. Summary of the Invention

[0006] In view of the above problems, this application provides a text steganography analysis method, apparatus, device, medium and program product.

[0007] According to a first aspect of this application, a text steganography analysis method is provided. The method includes, in response to receiving text to be detected, inputting a first feature vector of the text to be detected into a semantic baseline extraction layer of an analysis model, extracting low-frequency semantic features and artifact dimension signal features from the first feature vector, fusing the two to obtain a second feature vector, wherein the first feature vector is obtained by processing the text to be detected, and the analysis model further includes an artifact feature amplification layer, a global representation extraction layer, and a classification head module; inputting the first feature vector into the artifact feature amplification layer to amplify the artifact dimension signal features in the first feature vector to obtain a third feature vector including the amplified artifact dimension signal features; inputting the second feature vector and the third feature vector into the global representation extraction layer for feature fusion to obtain a fourth feature vector; inputting the fourth feature vector into the classification head module to obtain the detection result, wherein the detection result includes whether the text to be detected is normal text or steganographic text.

[0008] According to an embodiment of this application, extracting low-frequency semantic features and artifact dimension signal features from the first feature vector includes: performing contextual semantic encoding on the first feature vector based on the network parameters of the semantic benchmark extraction layer to obtain low-frequency semantic features; performing a lexical-level likelihood test on the lexical units in the text to be detected based on the first feature vector to determine the deviation value between the lexical occurrence probability and the statistical distribution of natural language, and encoding the deviation value to obtain statistical artifact feature components; determining structural anomaly information based on the content attention anomaly value and position attention anomaly value of the lexical units in the text to be detected, and encoding the structural anomaly information to obtain structural artifact feature components; and obtaining artifact dimension signal features based on the statistical artifact feature components and the structural artifact feature components.

[0009] According to an embodiment of this application, inputting the first feature vector into the artifact feature amplification layer and amplifying the artifact dimension signal features in the first feature vector to obtain a third feature vector including the amplified artifact dimension signal features includes: amplifying the low-frequency semantic features according to a preset low-rank adapter in the artifact feature amplification layer to obtain amplified low-frequency semantic features; amplifying the artifact dimension signal features according to the low-rank adapter to obtain amplified artifact dimension signal features; and obtaining the third feature vector based on the amplified low-frequency semantic features and the amplified artifact dimension signal features.

[0010] According to embodiments of this application, the low-rank adapter includes a multiplied dimensionality-reduced matrix and an dimensionality-increasing matrix.

[0011] According to an embodiment of this application, the process of inputting the second feature vector and the third feature vector into the global representation extraction layer for feature fusion to obtain a fourth feature vector includes: adding the low-frequency semantic features in the second feature vector to the amplified low-frequency semantic features in the third feature vector to obtain a summed low-frequency semantic feature; adding the artifact dimension signal features in the second feature vector to the amplified artifact dimension signal features in the third feature vector to obtain a summed artifact dimension signal feature; and obtaining the fourth feature vector based on the summed low-frequency semantic feature and the summed artifact dimension signal feature.

[0012] According to an embodiment of this application, inputting the fourth feature vector into the classification head module to obtain the detection result includes: determining the probability value of the text to be detected as normal text or steganographic text based on the summed artifact dimension signal features by the classification head module; and determining the detection result based on the probability value of the text to be detected as normal text or steganographic text.

[0013] A second aspect of this application provides a text steganography analysis apparatus, comprising: a first feature extraction module, configured to, in response to receiving text to be detected, input a first feature vector of the text to be detected into a semantic baseline extraction layer of an analysis model, extract low-frequency semantic features and artifact dimension signal features from the first feature vector, and fuse the two to obtain a second feature vector, wherein the first feature vector is obtained by processing the text to be detected, and the analysis model further comprises an artifact feature amplification layer, a global representation extraction layer, and a classification head module; a second feature extraction module, configured to input the first feature vector into the artifact feature amplification layer, amplify the artifact dimension signal features in the first feature vector, and obtain a third feature vector including the amplified artifact dimension signal features; a feature fusion module, configured to input the second feature vector and the third feature vector into the global representation extraction layer for feature fusion to obtain a fourth feature vector; and a text detection module, configured to input the fourth feature vector into the classification head module to obtain the detection result, wherein the detection result includes whether the text to be detected is normal text or steganographic text.

[0014] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0015] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0016] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method. Attached Figure Description

[0017] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0018] Figure 1 A flowchart illustrating a text steganography analysis method according to an embodiment of this application is shown schematically.

[0019] Figure 2 A schematic diagram of an analysis model according to an embodiment of this application is shown.

[0020] Figure 3 The illustration shows a schematic diagram of the process of training an analysis model according to an embodiment of this application;

[0021] Figure 4 A schematic block diagram of a text steganography analysis apparatus according to an embodiment of this application is shown. Detailed Implementation

[0022] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0023] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0024] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0025] Figure 1 A flowchart illustrating a text steganography analysis method according to an embodiment of this application is shown schematically; as follows: Figure 1As shown, step 110 is first executed. In response to receiving the text to be detected, the first feature vector of the text to be detected is input into the semantic baseline extraction layer of the analysis model. Low-frequency semantic features and artifact dimension signal features are extracted from the first feature vector, and the two are fused to obtain the second feature vector. The first feature vector is obtained by processing the text to be detected. The analysis model also includes an artifact feature amplification layer, a global representation extraction layer, and a classification head module. Extracting low-frequency semantic features and artifact dimension signal features from the first feature vector includes: performing contextual semantic encoding on the first feature vector according to the network parameters of the semantic baseline extraction layer to obtain low-frequency semantic features; performing a word-level likelihood test on the words in the text to be detected based on the first feature vector to determine the deviation value between the word occurrence probability and the statistical distribution of natural language, and encoding the deviation value to obtain statistical artifact feature components; determining structural anomaly information based on the content attention anomaly value and position attention anomaly value of the words in the text to be detected, and encoding the structural anomaly information to obtain structural artifact feature components; and obtaining artifact dimension signal features based on the statistical artifact feature components and the structural artifact feature components.

[0026] According to one implementation, after receiving the text to be detected, it undergoes preprocessing. This preprocessing is a fundamental step in steganalysis detection, primarily converting the unstructured text into a structured numerical feature vector recognizable by the model. First, it receives text sequences from multiple channels such as social media, instant messaging tools, and email. These sequences are raw, unprocessed natural language text. Then, it uses the same word segmenter as in the training phase to segment the text, mapping the continuous text sequence into a discrete token ID sequence, thus converting natural language into machine-recognizable discrete tokens. Next, a special classification marker (CLS) is forcibly inserted at the beginning of the token ID sequence. This marker is used for subsequent extraction of global aggregate features of the text. Simultaneously, a termination marker (SEP) is inserted at the end of the sequence to define the effective boundaries of the text sequence, allowing the model to accurately identify the start and end positions of the text. Next, based on the model's preset input specifications, the token ID sequence after adding special markers is padded or truncated. If the sequence length exceeds a preset value, it is truncated; if it is shorter than the preset value, it is padded with zeros. Finally, a standardized input tensor of shape [1, L] is generated, where L is the model's preset fixed sequence length to ensure uniform feature dimensions in the input model. Finally, this standardized token ID sequence is input into the model's word embedding layer. Through the vector mapping operation of the word embedding layer, the discrete token IDs are transformed into low-dimensional dense vector representations that incorporate word semantic information. This vector serves as the input basis for subsequent feature extraction in the model, realizing the transformation from discrete identifiers to continuous numerical features, allowing the model to perform subsequent steganography artifact feature extraction and discrimination operations based on this vector.

[0027] According to one implementation, the first feature vector of the text to be detected is then input into the semantic benchmark extraction layer of the analysis model. This layer performs feature extraction based on the network parameters W0 that are fixed and frozen after training. It simultaneously extracts low-frequency semantic features and artifact dimension signal features from the first feature vector. The two types of features are then deeply integrated into a fused feature vector as the second feature vector output. The first feature vector is obtained by standardizing preprocessing operations such as word segmentation and word embedding from the text to be detected. It is a low-dimensional dense vector representation that integrates the semantic information of all words in the text to be detected. This semantic benchmark extraction layer is the core foundational layer for the analysis model to achieve steganalysis detection. The second feature vector output by this layer provides a complete and reliable feature foundation for the subsequent artifact feature amplification layer, global representation extraction layer, and classification head module.

[0028] Before performing feature extraction, the relevant terms are explained: A lexical unit is the smallest semantic unit processed by a natural language processing model. It is the basic processing unit obtained after word segmentation and can include complete words, word fragments, punctuation marks, and other semantic carriers. Low-frequency semantic features are the core macroscopic semantic representations of the text, providing stable information that reflects the contextual logic and core semantic meaning of the text to be detected, and do not change with subtle local features. Artifact dimension signal features are the concentrated feature representations of steganography artifacts in the text to be detected, including statistical artifact feature components and structural artifact feature components. These are the extremely weak feature perturbations brought to the text when the steganography algorithm embeds secret information. Statistical artifact feature components represent the feature components that characterize statistical distribution shift artifacts in the text to be detected, caused by probability distribution modification steganography operations. Structural artifact feature components represent the feature components that characterize structural misalignment artifacts in the text to be detected. This is triggered by structural constraint-based steganography operations; lexical-level likelihood testing calculates the probability of each lexical occurrence in the current text context and compares this probability with the inherent statistical distribution of natural language to obtain a deviation value; content attention is a quantitative indicator representing the degree of semantic association between lexical units in the text, reflecting the interdependence of lexical units at the semantic level; positional attention is a quantitative indicator representing the degree of logical matching between the positional dimension of lexical units in the text and the context, reflecting the fit between the position of lexical units and the overall text context; the replacement word detection (RTD) mechanism is a label-level likelihood ratio testing mechanism inherent in pre-trained language models, which naturally has the ability to distinguish between original lexical units and anomalous lexical units replaced by generative sampling; the decoupled attention (DA) mechanism is an attention calculation mechanism inherent in pre-trained language models, which can decouple content attention and positional attention into independent calculation branches, realizing separate modeling of semantic association and positional association of lexical units.

[0029] The semantic baseline extraction layer first performs contextual semantic encoding on the first feature vector based on the frozen network parameters W0. By performing deep modeling on all the words contained in the first feature vector, it mines the contextual relationships between words, the semantic logic and expressive meaning of the text as a whole, and then extracts the low-frequency semantic features of the text to be detected. These low-frequency semantic features can stably present the macroscopic semantic background of the text to be detected and will not change due to the subtle steganalytic perturbations in the text. This provides a stable and reliable semantic reference baseline for the subsequent identification and differentiation of steganalytic artifacts, ensuring that the subsequent operations can use this semantic baseline as a reference to accurately identify abnormal artifact signals that deviate from the baseline.

[0030] While extracting low-frequency semantic features, the semantic benchmark extraction layer relies on its built-in replacement word detection (RTD) mechanism to perform word-level likelihood tests on each word in the text to be detected. First, it calculates the actual occurrence probability of each word in the current text context based on the network parameter W0. Then, it retrieves the natural language inherent statistical distribution learned in the pre-training stage and compares the actual occurrence probability of the word with the expected occurrence probability of the word in the corresponding context in the natural language statistical distribution one by one. The deviation value between the two is calculated. This deviation value can accurately reflect the degree of deviation between the occurrence probability of the word and the natural language distribution. Subsequently, the deviation value is processed by feature encoding and transformed into a numerical feature that can be processed by the model. This is the statistical artifact feature component. This component can accurately characterize the statistical distribution shift artifacts in the text to be detected caused by the truncation sampling and other operations of probability distribution modification steganography algorithms such as AC and Discop. It is one of the core components of the artifact dimension signal feature.

[0031] Subsequently, the semantic baseline extraction layer, relying on its built-in decoupled attention (DA) mechanism, decouples the attention features of each word in the text to be detected, separating the originally fused attention features into independent content attention features and positional attention features. Then, based on the network parameter W0, outlier calculations are performed on the decoupled content attention features and positional attention features. By comparing these outliers with the normal content attention and positional attention distributions in natural language, content attention outliers and positional attention outliers for each word are obtained. Content attention outliers quantify the unnaturalness of semantic associations between words, while positional attention outliers quantify the deviation between word position and contextual logic. The content attention and positional attention outliers of all words are then integrated and analyzed (e.g., summation). Combining the distribution characteristics and numerical magnitudes of the two types of outliers, structural anomalies in the text to be detected are identified. These structural anomalies are caused by VS... The structural artifact feature component is obtained by imposing global constraints on structural constraint steganography algorithms, which results in features such as incoherent local context, misaligned context logic, and unreasonable word collocation. This structural anomaly information is then processed by feature encoding, which transforms it into numerical features that can be processed by the model. This results in the structural artifact feature component, which can accurately characterize structural misalignment artifacts in the text to be detected and is another core component of the artifact dimension signal feature.

[0032] After obtaining the statistical artifact feature components and the structural artifact feature components, the semantic benchmark extraction layer performs feature fusion processing on the two types of feature components. By concatenating or adding elements one by one, the two are integrated into a complete feature vector, which is the artifact dimension signal feature of the text to be detected. This feature is a concentrated representation of all steganalytic artifacts in the text to be detected. If the text to be detected is normal natural text, all values ​​of its artifact dimension signal feature are approximately 0. If the text to be detected is steganalytic text, its artifact dimension signal feature is a non-zero weak value δ. This weak value δ is the extremely weak steganalytic perturbation introduced by the steganography algorithm when embedding secret information in the text. It is the core feature basis for distinguishing normal text from steganalytic text.

[0033] Finally, the semantic benchmark extraction layer deeply integrates the low-frequency semantic features and artifact dimension signal features extracted in the previous stage. Through feature fusion operation, the two types of features are fused into a unified fused feature vector, which is the second feature vector. It contains both low-frequency semantic features that can stably present the core semantics of the text, providing semantic anchors for subsequent operations, and artifact dimension signal features that can characterize text steganalysis artifacts, providing core discrimination criteria for subsequent steganalysis detection.

[0034] The core function of the semantic baseline extraction layer is to provide a stable and reliable semantic anchor for the entire text steganalysis process. Relying on the frozen network parameters W0, it ensures the stability, consistency, and accuracy of low-frequency semantic feature extraction and artifact dimension signal feature perception, avoiding semantic representation deviation and artifact detection errors caused by parameter updates. At the same time, without introducing any additional modules, adding any additional inputs, or changing any network structure, this layer uses its built-in RTD and DA mechanisms to achieve synchronous extraction and deep integration of low-frequency semantic features and artifact dimension signal features. This allows the subsequent artifact feature amplification layer to accurately amplify steganalysis artifacts based on the second feature vector output by this layer, while preserving the stable semantic baseline. This effectively avoids the model overfitting to the semantic content of the text in subsequent analysis, ensuring that steganalysis artifacts always exist in the form of abnormal signals relative to the semantic baseline. This provides a reliable feature foundation containing complete semantic background and accurate steganalysis artifact information for the feature aggregation of the subsequent global representation extraction layer and the steganalysis determination stage of the classification head module.

[0035] Subsequently, step 120 is executed, whereby the first feature vector is input into the artifact feature amplification layer to amplify the artifact dimension signal features in the first feature vector, resulting in a third feature vector including the amplified artifact dimension signal features. This includes: amplifying low-frequency semantic features using a preset low-rank adapter in the artifact feature amplification layer to obtain amplified low-frequency semantic features; amplifying artifact dimension signal features using the low-rank adapter to obtain amplified artifact dimension signal features; and obtaining the third feature vector based on the amplified low-frequency semantic features and the amplified artifact dimension signal features. The low-rank adapter includes a multiplicative dimension reduction matrix and a dimension increase matrix.

[0036] After the first feature vector is input into the artifact feature amplification layer, this layer uses a preset low-rank adapter as the core processing unit to amplify the low-frequency semantic features and artifact dimensionality signal features contained in the first feature vector, respectively. Then, the two types of amplified features are integrated to finally output a third feature vector containing the amplified artifact dimensionality signal features. The low-rank adapter is the core functional module of the artifact feature amplification layer, which is obtained by multiplying the dimension reduction matrix A and the dimension increase matrix B. That is, the matrix representation of the low-rank adapter is BA, and the dimension reduction matrix A∈R r×k The upgraded matrix B∈R d×r The rank r is much smaller than the feature dimension of the model itself. Relying on this low-rank structure, we can achieve efficient amplification of features and significantly reduce computational overhead.

[0037] The artifact feature amplification layer is a functional layer in the analysis model specifically designed for signal amplification of steganalytic artifact features. Its core function is to directionally amplify the extremely weak artifact dimension signal features in the first feature vector, making them stand out from the semantic features. The low-rank adapter is a feature amplification module built based on LoRA low-rank adaptation technology. It consists of a combination of dimensionality reduction and dimensionality increase matrices. It amplifies the target features through low-rank matrix transformations and uses very few trainable parameters. The dimensionality reduction matrix is ​​used in the low-rank adapter to map high-dimensional features to a low-dimensional feature space, achieving feature dimension compression. The dimensionality increase matrix is ​​used in the low-rank adapter to restore the features from the low-dimensional feature space to the original high-dimensional feature space. The matrix, in conjunction with the dimensionality reduction matrix, completes the low-rank transformation and amplification of features. The amplified low-frequency semantic features are the feature representations obtained after the low-frequency semantic features in the first feature vector are transformed by the low-rank adapter matrix, providing a benchmark for subsequent feature fusion. The amplified artifact dimension signal features are the feature representations obtained after the extremely weak artifact dimension signal features in the first feature vector are transformed by the low-rank adapter matrix. They are the core features that distinguish normal text from stegtext, and after amplification, they have feature amplitudes that can be recognized by the model. The third feature vector is the feature vector obtained by integrating the amplified low-frequency semantic features and the amplified artifact dimension signal features, providing a basis for the model's subsequent global representation extraction that includes the amplified artifact features.

[0038] The artifact feature amplification layer first calls a preset low-rank adapter to perform matrix transformation on the low-frequency semantic features in the first feature vector. The low-frequency semantic features are then input into the low-rank adapter, which consists of a dimension reduction matrix and an increase matrix. The low-frequency semantic features are then amplified by the low-dimensional projection of the dimension reduction matrix and the high-dimensional recovery of the increase matrix. This process yields amplified low-frequency semantic features that retain the macroscopic semantic baseline attributes of the original low-frequency semantic features, providing a semantic reference for the amplified artifact dimensional signal features.

[0039] While amplifying low-frequency semantic features, the artifact feature amplification layer continues to rely on the low-rank adapter to perform the same matrix transformation operation on the artifact dimension signal features in the first feature vector. The extremely weak artifact dimension signal features are input into the low-rank adapter, first mapped to the low-rank feature space by a dimension reduction matrix, and then restored to the original feature dimension by an dimension increase matrix. Through this low-rank transformation process, the artifact dimension signal features are amplified in a targeted manner, resulting in amplified artifact dimension signal features. The extremely weak steganalysis δ in the original steganographic text is transformed into a significant feature signal after this amplification process, greatly improving the recognizability of artifact features. Meanwhile, the artifact dimension signal features that are approximately 0 in the normal text remain approximately 0 after amplification, thereby achieving feature differentiation between normal text and steganographic text.

[0040] After obtaining the amplified low-frequency semantic features and the amplified artifact dimensional signal features, the artifact feature amplification layer performs feature fusion processing on the two types of features. The resulting feature vector is the third feature vector, which includes both the amplified low-frequency semantic features that serve as the semantic benchmark and the artifact dimensional signal features that have been directionally amplified and have significant feature amplitudes. This effectively highlights the artifact features from the semantic features, providing a feature foundation containing clear and significant steganalysis artifact features for the feature aggregation of the subsequent global representation extraction layer and the steganalysis determination of the classification head module.

[0041] The core function of the artifact feature amplification layer is to achieve targeted amplification of the artifact dimension signal features by relying on the low-rank adapter. Through low-rank matrix transformation, without increasing too much computational overhead, the extremely weak and difficult-to-identify steganatomical artifact features in the first feature vector are amplified to become significant and identifiable features. At the same time, low-frequency semantic features are amplified synchronously to retain a stable semantic benchmark, ensuring that the artifact features always exist in the form of abnormal signals relative to the semantic benchmark. This effectively improves the model's ability to identify steganatomical artifacts and provides accurate and reliable feature basis for subsequent steganatomical detection. Furthermore, the low-parameter characteristics of the low-rank adapter significantly reduce the model's computational overhead and memory usage, ensuring the efficiency of steganatomical detection.

[0042] Subsequently, step 130 is executed, in which the second feature vector and the third feature vector are input into the global representation extraction layer for feature fusion to obtain the fourth feature vector; including: adding the low-frequency semantic features in the second feature vector to the amplified low-frequency semantic features in the third feature vector to obtain the summed low-frequency semantic features; adding the artifact dimension signal features in the second feature vector to the amplified artifact dimension signal features in the third feature vector to obtain the summed artifact dimension signal features; and obtaining the fourth feature vector based on the summed low-frequency semantic features and the summed artifact dimension signal features.

[0043] According to one implementation, after completing the feature processing of the semantic benchmark extraction layer and the artifact feature amplification layer, the second feature vector and the third feature vector are input into the global representation extraction layer. This layer performs feature fusion operations on the corresponding feature components of the two types of feature vectors respectively. After component-level summation processing, the fourth feature vector is obtained by integrating the summed low-frequency semantic features and the summed artifact dimension signal features. This vector is a global feature representation that covers the complete semantic background of the text and significant steganalysis artifact information, providing core feature basis for steganalysis determination in the subsequent classification head module.

[0044] The global representation extraction layer is a functional layer in the analysis model that realizes feature fusion and global representation construction. Its core function is to fuse the basic features output by the semantic baseline extraction layer and the amplified features output by the artifact feature amplification layer at the component level to construct a global feature vector containing complete text feature information. The fourth feature vector is the global feature vector output by the global representation extraction layer after fusing the second and third feature vectors. It contains both the summed low-frequency semantic features and the summed artifact dimension signal features, integrating all the information of the basic features and the amplified features. The summed low-frequency semantic features are the feature representation obtained by adding the low-frequency semantic features of the second feature vector and the amplified low-frequency semantic features of the third feature vector element by element. It is the semantic baseline feature of the text globally. The summed artifact dimension signal features are the feature representation obtained by adding the artifact dimension signal features of the second feature vector and the amplified artifact dimension signal features of the third feature vector element by element. It is the steganalytic artifact feature of the text globally. It integrates the basic artifact features and the amplified artifact features, further improving the recognizability of the artifact features.

[0045] The global representation extraction layer first fuses the corresponding low-frequency semantic feature components in the second and third feature vectors. It performs element-wise addition of the unamplified low-frequency semantic features in the second feature vector and the amplified low-frequency semantic features in the third feature vector after low-rank adapter, resulting in a summed low-frequency semantic feature. This feature strengthens the semantic baseline feature representation while preserving the original macro-semantic background of the text, ensuring that the subsequent determination of steganalysis artifacts always has a stable and unified semantic reference. At the same time, it ensures the integrity and consistency of semantic features and avoids the loss of semantic information due to feature fusion.

[0046] Subsequently, the global representation extraction layer fuses the artifact dimension signal feature components corresponding to the second and third feature vectors. It performs element-wise addition of the extremely weak original artifact dimension signal feature in the second feature vector and the artifact dimension signal feature in the third feature vector after being amplified by the low-rank adapter, to obtain the summed artifact dimension signal feature. This feature superimposes the original weak steganalytic perturbation δ in the steganographic text with the amplified artifact signal, further amplifying the amplitude of the artifact feature and making the steganographic artifacts stand out more significantly from the semantic features. Meanwhile, the artifact dimension signal feature, which is approximately 0 in the normal text, remains approximately 0 after summing, thereby further strengthening the feature difference between the normal text and the steganographic text and improving the model's ability to recognize steganographic artifacts.

[0047] After obtaining the summed low-frequency semantic features and the summed artifact dimension signal features, the global representation extraction layer integrates the two types of summed feature components. By concatenating features, the summed low-frequency semantic features and the summed artifact dimension signal features are fused into a unified high-dimensional feature vector, which is the fourth feature vector. It contains both the summed low-frequency semantic features, which serve as the global semantic benchmark and can stably present the core semantic background of the text to be detected, and the summed artifact dimension signal features, which have been doubly enhanced and have significant feature amplitudes. This can accurately characterize whether steganalysis artifacts exist in the text to be detected and the significance of the artifacts. This achieves a deep fusion of basic features and amplified features, and constructs a global feature representation of the text to be detected.

[0048] The core function of the global representation extraction layer is to achieve precise fusion of basic features and amplified features at the component level. By summing low-frequency semantic features and artifact dimension signal features respectively, it not only preserves and strengthens the semantic baseline features of the text, but also further amplifies and highlights the steganalysis artifact features. This effectively avoids the semantic features from masking the artifact features during the feature fusion process, so that the fused fourth feature vector has both a stable semantic background and significant artifact discrimination features. This provides a complete and accurate global feature basis for the subsequent steganalysis probability calculation and detection result determination of the classification head module.

[0049] Finally, step 140 is executed, the fourth feature vector is input into the classification head module, and the detection result is obtained. The detection result includes whether the text to be detected is normal text or steganographic text; including: determining the probability value of the text to be detected as normal text or steganographic text based on the summed artifact dimension signal features by the classification head module; and determining the detection result based on the probability value of the text to be detected as normal text or steganographic text.

[0050] After the feature fusion of the global representation extraction layer yields the fourth feature vector, this vector is input into the classification head module. This module uses the summation artifact dimension signal features in the fourth feature vector as the core discrimination criterion, and completes the final determination of steganalysis detection through feature space transformation and probability mapping, outputting the detection result of whether the text to be detected is normal text or steganalysis. The core technical terms involved in this step are explained as follows: The classification head module is the functional module in the text steganalysis model that realizes the final binary classification determination. It consists of a linear projection layer and a nonlinear activation function. Its function is to map the high-dimensional fourth feature vector into a scalar probability value between 0 and 1, and to determine whether the text is steganalysis or normal text based on the probability value. The probability value is a scalar score output by the classification head module representing the confidence that the text to be detected belongs to steganalysis, with a value range of [0,1]. The higher the score, the more significant the statistical distribution shift and structural misalignment in the text to be detected, i.e., the greater the probability that it belongs to steganalysis. The detection result is the binary classification determination result output by the classification head module based on the probability value and a preset threshold, containing only two results: the text to be detected is normal text or steganalysis.

[0051] After receiving the fourth feature vector, the classification head module first extracts the summation artifact dimension signal feature from it. This feature is a global comprehensive representation of the steganalysis artifacts in the text to be detected. The summation artifact dimension signal feature value of normal text is approximately 0, while that of steganalysis text is a significantly non-zero value after being amplified twice. The classification head module uses this feature as the basis for judgment throughout the process. Subsequently, the classification head module performs feature space compression on the summation artifact dimension signal feature through a built-in linear projection layer, mapping the high-dimensional artifact feature vector from the original feature space to the binary classification discriminant space. This achieves dimensionality reduction of high-dimensional features and extraction of discriminant features, retaining the core information in the summation artifact dimension signal feature used to distinguish between normal text and steganalysis text.

[0052] After linear projection, the classification head module performs a nonlinear mapping on the projected features using a nonlinear activation function, mapping the feature values ​​from the real number domain to the probability domain [0,1], generating a probability value for whether the text to be detected belongs to steganography. This probability value directly reflects the significance of the summed artifact dimension signal features. If the text to be detected is normal text, its summed artifact dimension signal features are approximately 0, and the probability value obtained after mapping is close to 0; if the text to be detected is steganography, its summed artifact dimension signal features are significantly non-zero, and the probability value obtained after mapping is close to 1. Simultaneously, this probability value can also represent the probability that the text to be detected is normal text, i.e., the probability value of normal text is 1 minus the probability value of steganography.

[0053] After obtaining the probability value of the text to be detected being steganographic, the classification head module compares this probability value with a preset fixed judgment threshold. This threshold is the optimal binary classification judgment critical value determined during the model training phase, typically preset to 0.5. If the probability value is greater than or equal to the preset threshold, it indicates that the summation artifact dimension signal feature of the text to be detected is significant, with obvious statistical distribution shift or structural misalignment, and the detection result is determined to be steganographic text. If the probability value is less than the preset threshold, it indicates that the summation artifact dimension signal feature of the text to be detected is approximately 0, with no obvious steganographic artifacts, and the detection result is determined to be normal text.

[0054] The core function of the classification head module is to map high-dimensional features to probability values ​​and make a final binary classification decision, based on the summation of artifact dimensional signal features. Through a combination of linear projection and nonlinear activation, it accurately extracts the discriminative information from the artifact features and transforms it into interpretable probability values. Then, it uses a preset threshold to achieve automated detection result determination. The entire process focuses only on the amplified and fused artifact features, removing the interference of semantic features to ensure the accuracy and robustness of the detection results. At the same time, the lightweight structure design of this module also ensures the computational efficiency of the steganalysis detection inference stage, which meets the deployment requirements of low overhead and high real-time performance of the model.

[0055] Figure 2 A schematic diagram of an analysis model according to an embodiment of this application is shown. Figure 2 As shown, the analysis model includes a lexical preprocessing and word embedding layer 210, a semantic benchmark extraction layer 220, an artifact feature amplification layer 230, a global representation extraction layer 240, and a classification head module 250. The lexical preprocessing and word embedding layer 210 is connected to the semantic benchmark extraction layer 220 and the artifact feature amplification layer 230; the global representation extraction layer 240 is connected to the semantic benchmark extraction layer 220 and the artifact feature amplification layer 230. The global representation extraction layer 240 is connected to the classification head module 250.

[0056] The input to the lexical preprocessing and word embedding layer 210 is the raw, unprocessed text sequence to be detected, which can originate from any channel of natural language text, such as social media, instant messaging tools, and emails. It is used to complete the standardized transformation from unstructured natural language to structured numerical features, following the preprocessing rules from the model training stage to ensure consistency of feature representation and avoid feature bias caused by differences in processing rules. Specifically, it performs two steps: lexicalization and word embedding mapping. First, the original text sequence is mapped into a discrete sequence of token IDs using a word segmenter. A special classification marker (CLS) is inserted at the beginning of the sequence, and a termination marker (SEP) is inserted at the end. The sequence is then padded or truncated according to the model's preset specifications to generate a standardized token ID sequence with a fixed shape. This sequence is then input into the word embedding layer, where vector mapping transforms the discrete token IDs into a low-dimensional dense vector that incorporates lexical semantic information. The output is a low-dimensional dense vector (i.e., the first feature vector) that incorporates lexical semantic information, providing a unified feature input basis for the semantic benchmark extraction layer 220 and the artifact feature amplification layer 230.

[0057] The semantic baseline extraction layer 220 performs feature extraction based on the fixed network parameters W0 frozen after training. It simultaneously extracts low-frequency semantic features and artifact-dimensional signal features, providing a stable semantic anchor for the entire model and achieving preliminary perception of steganalysis artifacts. This process avoids introducing additional modules or updating parameters, thus preventing semantic overfitting. Specific operations include: performing contextual semantic encoding on the input feature vector based on network parameters W0 to mine the relationships between word units and the overall semantic logic of the text, extracting low-frequency semantic features that stably present the macro-level semantic background of the text; relying on the built-in replacement word detection (RTD) mechanism, performing word-level likelihood testing on each word unit, calculating the deviation between the word unit occurrence probability and the natural language statistical distribution, and encoding it as a statistical artifact feature component; relying on the built-in decoupled attention (DA) mechanism, separating the content attention and positional attention of word units, calculating the outliers of the two types of attention, and integrating and encoding them as a structural artifact feature component; fusing the statistical and structural artifact feature components into artifact-dimensional signal features, and finally integrating the low-frequency semantic features and artifact-dimensional signal features into a unified feature vector. The output of the semantic baseline extraction layer 220 contains a fused feature vector (i.e., the second feature vector) that combines low-frequency semantic features and artifact dimension signal features. The artifact dimension signal features of normal text are approximately 0, while the artifact dimension signal features of steganographic text are non-zero weak values ​​δ.

[0058] The artifact feature amplification layer 230 uses a low-rank adapter built based on LoRA technology as its core processing unit to achieve targeted amplification of artifact-dimensional signal features in the feature vector, while simultaneously amplifying low-frequency semantic features. While preserving the semantic baseline, it transforms extremely weak steganalytic artifact features into salient features recognizable by the model. The low-rank adapter consists only of a dimensionality reduction matrix A and an dimensionality increase matrix B, with very few parameters, ensuring model detection efficiency. Specific operations include: separating low-frequency semantic features from artifact-dimensional signal features in the input feature vector; sequentially projecting the low-rank adapter's dimensionality reduction matrix into a low-dimensional form and restoring it into a high-dimensional form; amplifying the low-frequency semantic features to obtain amplified low-frequency semantic features; and targeted amplifying the artifact-dimensional signal features to obtain amplified artifact-dimensional signal features (the weak perturbation δ of the steganalytic text is significantly amplified); and integrating the two types of amplified features into a unified feature vector. The artifact feature amplification layer 230 outputs a fused feature vector (i.e., the third feature vector) that simultaneously contains amplified low-frequency semantic features and amplified artifact dimensional signal features, thus achieving the initial highlighting of steganalytic artifact features from semantic features.

[0059] The input to the global representation extraction layer 240 includes the second feature vector output by the semantic baseline extraction layer 220 and the third feature vector output by the artifact feature amplification layer 230. The global representation extraction layer 240 is the core layer for feature fusion of the model. It performs component-level precise fusion on the feature vectors output by the two parallel processing layers, deeply integrating basic features and amplified features to construct a global discriminative feature that covers the complete semantic background of the text and significant steganalysis artifact information. This provides a comprehensive and accurate feature basis for subsequent judgment, while avoiding the masking of artifact features by semantic features during feature fusion. Specific operations include: adding the low-frequency semantic features in the second feature vector to the amplified low-frequency semantic features in the third feature vector element-wise to obtain a summed low-frequency semantic feature, strengthening the feature representation of the semantic baseline; adding the artifact dimension signal features in the second feature vector to the amplified artifact dimension signal features in the third feature vector element-wise to obtain a summed artifact dimension signal feature, further amplifying the artifact feature amplitude and strengthening the feature differences between normal or steganalysis text; and integrating the summed low-frequency semantic features and the summed artifact dimension signal features into a unified global feature vector. The global representation extraction layer 240 outputs a global fusion feature vector (i.e., the fourth feature vector) that simultaneously contains the summed low-frequency semantic features and the summed artifact dimension signal features. This vector is the core discriminant feature that covers the global distribution anomalies and structural anomalies of the text.

[0060] The classification head module 250 takes as input the fourth feature vector output by the global representation extraction layer 240. As the final judgment layer of the model, it uses the summation artifact dimension signal feature in the fourth feature vector as the sole core discrimination criterion, removing interference from low-frequency semantic features and completing the conversion from high-dimensional global features to binary classification detection results, thus achieving automated and accurate judgment between stegtext and normal text. Specific operations include: extracting the summation artifact dimension signal feature from the fourth feature vector; compressing the high-dimensional artifact feature space to the binary classification discrimination space through a linear projection layer, retaining core discrimination information; mapping the projected feature values ​​from the real number domain to the [0,1] probability domain through a non-linear activation function, outputting the probability value that the text to be detected belongs to stegtext (a higher score indicates a more significant distribution shift and structural misalignment in the text); comparing the probability value with a preset fixed threshold (usually 0.5), and outputting the final detection result based on the comparison result. The classification head module 250 outputs the binary classification detection results of the text to be detected, which includes only two judgment results: if the probability value exceeds the preset threshold, it is judged as steganographic text; if the probability value is lower than the preset threshold, it is judged as normal text.

[0061] Figure 3 The illustration schematically depicts a process diagram for training an analysis model according to an embodiment of this application. For example... Figure 3 As shown, the first step is to load the training dataset and labels. This step is the basic data preparation stage for model training. The loaded dataset is a balanced label dataset across semantic domains. This dataset is constructed through steps such as multi-source corpus acquisition, cleaning, and steganalysis. Specifically, original text corpora are obtained from different semantic domains such as movie reviews, news reports, and social media. After data cleaning to remove format noise and sentence segmentation, text sequences of a preset length are retained. Then, mainstream generative steganography algorithms such as AC, Discop, and VS are used to steganographically embed some of the carrier text. Finally, a balanced number of labeled sample pairs of natural text (negative samples) and steganalytic text (positive samples) are constructed. The dataset provides diverse supervised learning samples for model training. The cross-domain corpus design lays the foundation for the model's subsequent cross-domain generalization ability, while the balanced positive and negative samples ensure that the model does not suffer from class bias and can learn the feature differences between normal text and steganalytic text in a balanced way.

[0062] First, initialize DeBERTa-v3 and freeze its parameters. This step completes the construction and parameter constraints of the semantic backbone network in the semantic baseline extraction layer of the model. The core is the pre-trained DeBERTa-v3 model, which serves as the semantic backbone network of the entire analysis model. This model has a bidirectional attention mechanism and natively includes a replacement word detection (RTD) pre-training task and a decoupled attention (DA) mechanism. It is highly sensitive to subtle features in the text and can accurately capture statistical distribution shifts and structural misalignment artifacts caused by steganography algorithms. At the same time, its parameter size is much smaller than that of mainstream large language models, which can effectively reduce the computational cost of training and inference. After completing the model initialization, set the weight parameters of all Transformer encoding layers of DeBERTa-v3 to a non-trainable frozen state. This transforms the pre-trained model into a semantic low-pass filter. Utilizing its stable representation capability of natural language manifold established in large-scale corpus pre-training, it achieves the isolation and extraction of low-frequency semantic information, avoiding the model from fitting the semantic content of the training data due to parameter updates during subsequent fine-tuning, thus fundamentally solving the semantic overfitting problem.

[0063] When injecting the LoRA adapter, a differential feature amplification module is constructed on the frozen semantic backbone network to obtain the artifact feature amplification layer. The core is in the self-attention layer and feedforward network layer of the DeBERTa-v3 backbone network. A low-rank adapter based on LoRA low-rank adaptation technology is injected in parallel. This adapter consists of a dimension reduction matrix A and an increase matrix B, where the rank r of the matrix is ​​much smaller than the feature dimension of the model itself. The whole acts as a differential feature amplifier (DFA). The low-rank adapter is the only core trainable module in the model. Its low-rank structure ensures the feature amplification effect while occupying very few trainable parameters, which greatly reduces the training cost and memory usage of the model. Injecting it into the core feature layer of the backbone network enables it to accurately capture the high-frequency residual signal orthogonal to the semantic manifold after filtering by the backbone network, which is the potential steganalysis artifact signal, providing a foundation for subsequent artifact differential amplification.

[0064] Subsequently, forward propagation extracts features. The loaded training samples are input into the model that has been initialized and module injected, and the entire feature forward propagation calculation is performed. Relying on the model's architecture alignment characteristics and differential amplification capabilities, the capture, extraction, and amplification of steganalysis artifact features are completed. During the propagation process, the native DA mechanism of DeBERTa-v3 is first used to separate the content vector and relative position vector of text words. Attention scores are calculated through three-way interaction to remove interference from strong semantic signals and accurately detect and capture structural artifact features such as local contextual incoherence and structural misalignment caused by algorithms such as VS. At the same time, the label-level likelihood ratio test capability of the RTD pre-training task is used to identify the statistical distribution shift caused by probability truncation in algorithms such as AC and Discop, and capture statistical artifact features. Subsequently, the injected LoRA low-rank adapter performs directional differential amplification on the captured high-frequency steganalysis signal, transforming the extremely weak steganalysis perturbation δ into a significant feature signal that the model can recognize. As for the low-frequency semantic information extracted by the backbone network, only task-adaptive processing is performed without excessive amplification. Finally, the feature stream that integrates the amplified distribution and structural features is injected into the classification head, providing a feature basis for loss calculation.

[0065] Next, the binary cross-entropy loss is calculated as the core basis for updating model parameters. First, the hidden state vector corresponding to the special classification label CLS in the last Transformer encoding block of the model is extracted. This vector is an aggregated feature that integrates global text features, covering the distribution and structural anomalies of the entire text sequence, and can serve as the core global feature for model discrimination. Then, this global aggregated feature is input into the classification head module. After feature projection, the model's predicted steganalysis probability for the sample is obtained. The predicted probability is then compared with the true binary classification label of the sample, and the binary cross-entropy loss value between the two is calculated. The binary cross-entropy loss is a classic loss function for binary classification tasks, which can accurately measure the degree of deviation between the model's prediction result and the true label. The higher the loss value, the lower the current steganalysis discrimination accuracy of the model, and vice versa. This loss value will serve as the core optimization target for subsequent model parameter updates.

[0066] Finally, the system performs model convergence determination, parameter iteration, and model saving. The model's training completion is determined by two dimensions: the convergence status of the loss function and the preset maximum number of iterations. The system first checks the convergence of the currently calculated binary cross-entropy loss value and simultaneously determines whether the current training iterations have reached the preset maximum number of iterations. If the loss function has converged, or the training iterations have reached their maximum value, it indicates that the model has fully learned the steganalysis artifact features. At this point, the system exits the training loop, outputs and saves the current optimal model parameters. The saved parameters include the optimized LoRA low-rank adapter parameters and the classification head parameters. The frozen DeBERTa-v3 backbone network parameters remain unchanged. The final result is a fully trained model with high-frequency artifact amplification capabilities and multi-dimensional architecture alignment characteristics, which can be directly used for real-time steganalysis in the subsequent inference and detection stages. If the loss function fails to converge and the training iterations have not reached their maximum value, it indicates that the model still has room for optimization. In this case, the AdamW optimizer is used to update the gradients of the only trainable parameters in the model (LoRA low-rank adapter parameters and classification head parameters). During the update process, the parameters of the DeBERTa-v3 backbone network are strictly kept frozen to ensure the stability of the semantic backbone and the semantic low-pass filtering effect. After the parameter update is completed, the model will return to the forward propagation feature extraction step, load the training dataset, and perform a new round of feature learning, loss calculation, and convergence determination, thus forming a training loop until the model meets the convergence condition or reaches the maximum number of iterations.

[0067] The entire training process freezes the semantic backbone and trains only the low-rank adapter. Through architecture alignment, the model accurately captures two types of steganalysis artifacts. Through differential feature amplification, the model focuses on learning high-frequency artifact signals. At the same time, relying on supervised optimization and parameter iteration using binary cross-entropy loss, the model achieves efficient recognition of steganalysis artifacts. The final trained model not only breaks through the detection accuracy bottleneck of traditional methods but also solves the computational efficiency problem of existing large-scale model methods, while possessing excellent cross-domain and cross-algorithm generalization capabilities.

[0068] This application effectively achieves a balance between high detection accuracy and high execution efficiency. By adopting a DeBERTa discriminative encoder architecture instead of the traditional Llama generative decoder architecture, only a small number of parameters need to be fine-tuned (the base parameters are about 0.4B, compared to the 7B+ large model base of existing technologies). While significantly reducing computational overhead, it achieves superior detection accuracy compared to existing large language model methods by relying on the precise alignment of RTD and DA mechanisms. It also reduces memory usage by 50% and improves training efficiency by 7 times. At the same time, it has extremely strong cross-domain and cross-algorithm generalization capabilities. By combining the strategy of freezing the semantic backbone with the differential feature amplification mechanism, the model can effectively suppress the interference of background semantic noise on the detection results, fundamentally avoiding overfitting to text content in specific domains. Even when facing unseen semantic domains or steganography algorithms, it can still maintain extremely high detection accuracy, demonstrating excellent cross-text domain and cross-algorithm domain generalization performance.

[0069] In experimental verification, the text steganalysis method of this invention showed excellent detection performance in three mainstream generative steganalysis algorithms (AC, DI, and VS) and in three different text domain combination test scenarios (Movie, News, and Twitter). It can effectively extract weak steganalytic features in text and comprehensively outperform existing mainstream detection methods such as TS-RNN, LSTMATT, and GS-Llama7b with a large model base in terms of the two core indicators of accuracy (ACC) and F1 score.

[0070] In a comparative experiment on computational resource consumption, using average training time and average GPU memory usage as the core evaluation metrics, this method was compared with the mainstream large language model steganalysis method GS-Llama7b. The results show that this invention significantly improves detection accuracy while greatly reducing computational overhead: Under the same hardware environment, the average training time of the standard version of this invention (Ours Std) is only 3.20 minutes, which is about 22 times faster than GS-Llama7b's 71.73 minutes, making it more suitable for the needs of real-time steganalysis; and even when the standard version of this invention uses a larger batch size (BatchSize = 12), its GPU memory usage is still much lower than that of GS-Llama7b with a smaller batch size. The GPU memory usage of the aligned version of this invention (OursAlign) is only 4.34GB, which is only about half that of GS-Llama7b.

[0071] Based on the above code testing method, this application also provides a text steganography analysis device. The following will combine... Figure 4 The device is described in detail. Figure 4 A schematic block diagram of a text steganography analysis apparatus according to an embodiment of this application is shown.

[0072] like Figure 4 As shown, the text steganography analysis device 400 of this embodiment includes a first feature extraction module 410, a second feature extraction module 420, a feature fusion module 430, and a text detection module 440.

[0073] The first feature extraction module 410 is used to respond to receiving the text to be detected by inputting the first feature vector of the text to be detected into the semantic baseline extraction layer of the analysis model, extracting low-frequency semantic features and artifact dimension signal features from the first feature vector, and using them as the second feature vector. The first feature vector is obtained by processing the text to be detected. The analysis model also includes an artifact feature amplification layer, a global representation extraction layer, and a classification head module. The first feature extraction module 410 can be used to perform step 110 described above, which will not be repeated here.

[0074] The second feature extraction module 420 is used to input the first feature vector into the artifact feature amplification layer, amplify the artifact dimension signal features in the first feature vector, and obtain a third feature vector including the amplified artifact dimension signal features; the second feature extraction module 420 can be used to execute step 120 described above, which will not be repeated here.

[0075] The feature fusion module 430 is used to input the second feature vector and the third feature vector into the global representation extraction layer to perform feature fusion to obtain the fourth feature vector; the feature fusion module 430 can be used to perform step 130 described above, which will not be repeated here.

[0076] The text detection module 440 is used to input the fourth feature vector into the classification head module to obtain the detection result, which includes whether the text to be detected is normal text or steganographic text; the text detection module 440 can be used to execute the step 140 described above, which will not be repeated here.

[0077] It should be noted that the storage medium (computer-readable medium) described above in this invention can be a computer-readable signal medium, a non-transitory computer-readable storage medium, or any combination thereof. A non-transitory computer-readable storage medium can be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a non-transitory computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0078] In this invention, a non-transitory computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a non-transitory computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination thereof.

[0079] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0080] The above description is merely a partial embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0081] Furthermore, although the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in sequential order. Multitasking and parallel processing may be advantageous in certain environments. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the invention. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0082] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A text steganography analysis method, characterized in that, The method includes: In response to receiving the text to be detected, the first feature vector of the text to be detected is input into the semantic baseline extraction layer of the analysis model. Low-frequency semantic features and artifact dimension signal features are extracted from the first feature vector, and the two are fused together as the second feature vector. The first feature vector is obtained by processing the text to be detected. The analysis model also includes an artifact feature amplification layer, a global representation extraction layer, and a classification head module. The first feature vector is input into the artifact feature amplification layer to amplify the artifact dimension signal features in the first feature vector, thereby obtaining a third feature vector that includes the amplified artifact dimension signal features. The second feature vector and the third feature vector are input into the global representation extraction layer for feature fusion to obtain the fourth feature vector; The fourth feature vector is input into the classification head module to obtain the detection result, which includes whether the text to be detected is normal text or steganographic text. Specifically, based on the first feature vector, a lexical-level likelihood test is performed on the lexical units in the text to be detected to determine the deviation value between the lexical occurrence probability and the statistical distribution of natural language, and the deviation value is encoded to obtain statistical artifact feature components. Structural anomaly information is determined based on the content attention anomaly value and position attention anomaly value of the word units in the text to be detected, and the structural anomaly information is encoded to obtain structural artifact feature components. The artifact dimension signal features are obtained based on the statistical artifact feature components and the structural artifact feature components.

2. The method as described in claim 1, characterized in that, Extracting low-frequency semantic features from the first feature vector includes: The first feature vector is subjected to contextual semantic encoding based on the network parameters of the semantic benchmark extraction layer to obtain low-frequency semantic features.

3. The method as described in claim 1, characterized in that, The first feature vector is input into the artifact feature amplification layer to amplify the artifact dimension signal features in the first feature vector, resulting in a third feature vector that includes the amplified artifact dimension signal features: The low-frequency semantic features are amplified by a preset low-rank adapter in the artifact feature amplification layer to obtain amplified low-frequency semantic features. The artifact dimension signal features are amplified using the low-rank adapter to obtain amplified artifact dimension signal features. The third feature vector is obtained based on the amplified low-frequency semantic features and the amplified artifact dimension signal features.

4. The method as described in claim 3, characterized in that, The low-rank adapter includes a multiplied dimensionality-reducing matrix and an dimensionality-increasing matrix.

5. The method according to any one of claims 1-4, characterized in that, The second feature vector and the third feature vector are input into the global representation extraction layer for feature fusion to obtain the fourth feature vector, which includes: The low-frequency semantic features in the second feature vector are added to the amplified low-frequency semantic features in the third feature vector to obtain the summed low-frequency semantic features; The artifact dimension signal feature in the second feature vector is added to the amplified artifact dimension signal feature in the third feature vector to obtain the summed artifact dimension signal feature. The fourth feature vector is obtained based on the summed low-frequency semantic features and the summed artifact dimension signal features.

6. The method as described in claim 5, characterized in that, The fourth feature vector is input into the classification head module to obtain the detection results, including: The classification head module determines the probability value of whether the text to be detected is normal text or steganographic text based on the summed artifact dimension signal features; The detection result is determined based on the probability value of whether the text to be detected is normal text or steganographic text.

7. A text steganography analysis device, characterized in that, The device includes: The first feature extraction module is used to respond to receiving the text to be detected by inputting the first feature vector of the text to be detected into the semantic baseline extraction layer of the analysis model, extracting low-frequency semantic features and artifact dimension signal features from the first feature vector, and fusing the two as the second feature vector. The first feature vector is obtained by processing the text to be detected. The analysis model also includes an artifact feature amplification layer, a global representation extraction layer, and a classification head module. The second feature extraction module is used to input the first feature vector into the artifact feature amplification layer, amplify the artifact dimension signal features in the first feature vector, and obtain a third feature vector including the amplified artifact dimension signal features. The feature fusion module is used to input the second feature vector and the third feature vector into the global representation extraction layer to fuse features and obtain a fourth feature vector; The text detection module is used to input the fourth feature vector into the classification head module to obtain the detection result, the detection result including whether the text to be detected is normal text or steganographic text; Specifically, based on the first feature vector, a lexical-level likelihood test is performed on the lexical units in the text to be detected to determine the deviation value between the lexical occurrence probability and the statistical distribution of natural language, and the deviation value is encoded to obtain statistical artifact feature components. Structural anomaly information is determined based on the content attention anomaly value and position attention anomaly value of the word units in the text to be detected, and the structural anomaly information is encoded to obtain structural artifact feature components. The artifact dimension signal features are obtained based on the statistical artifact feature components and the structural artifact feature components.

8. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1-6.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1-6.

Images