A zero-watermark generation and detection method based on multi-dimensional text features

Abstract

The application discloses a zero watermark generation and detection method based on multi-dimensional text features, and belongs to the technical field of natural language processing, which comprises the following steps: preprocessing and semantic normalization of input text to obtain normalized text; extracting multi-dimensional text features from the normalized text to construct zero watermark; generating zero watermark identification associated with the multi-dimensional text features, and storing the zero watermark identification; during copyright verification, extracting multi-dimensional text features from the text to be detected, respectively calculating the similarity between each dimension feature of the text to be detected and the stored zero watermark identification; constructing the similarity between each dimension feature into a joint feature vector, inputting the joint feature vector into a classification model for weighted fusion judgment, and outputting the copyright authentication result of the text to be detected. The application improves the robustness and accuracy of text detection, enhances the ability to resist structural attacks, and realizes effective defense against generative rewriting attacks.

Description

Technical Field

[0001] This invention belongs to the technical field of natural language processing technology, specifically relating to a method for generating and detecting zero watermarks based on multi-dimensional text features. Background Technology

[0002] With the rapid development of information technology and the widespread application of the internet, massive amounts of text data are growing exponentially. Whether it's academic research, news reports, or corporate documents, digital documents face the problem of being illegally tampered with, stolen, or maliciously modified during dissemination, causing serious damage to the interests of the original authors. Therefore, how to effectively protect text copyright and detect tampering has become a crucial issue that urgently needs to be addressed.

[0003] Zero-watermarking technology does not modify the original text. Instead, it extracts key features from the text itself to construct watermark information and registers it with a third-party certification authority or on a blockchain. Because it maintains the integrity of the text, zero-watermarking technology has become a research hotspot in text copyright protection.

[0004] However, existing methods for constructing zero-watermark text still have the following significant shortcomings when dealing with complex attack methods: First, methods based on statistical features lack robustness. Many existing methods rely primarily on statistical features such as word frequency, sentence length, or punctuation distribution. While these features are computationally simple, they lack semantic understanding. Attackers can significantly alter these statistical features without changing the original meaning by simply replacing synonyms (e.g., changing "beautiful" to "pretty") or making minor adjustments to sentence structure, causing watermark detection to fail.

[0005] Second, single-level semantic features struggle to capture both details and the overall picture. Some deep learning-based methods attempt to extract semantic features using Word2Vec or earlier BERT models. However, these methods typically either focus too much on local words, making them sensitive to changes in the text structure, or focus too much on the overall vector, making them insensitive to alterations to key numbers or entities. For example, methods relying solely on the entire text vector would struggle to detect subtle but critical modifications, such as changing "compensation of 1 million yuan" to "compensation of 100,000 yuan."

[0006] Third, there is a lack of defense against structural attacks and generative AI plagiarism. When faced with structural attacks such as paragraph rearrangement and sentence reordering, sequence-based feature extraction methods are prone to misalignment. More seriously, with the widespread adoption of large language models like ChatGPT, attackers can easily perform deep rewriting of text, i.e., "plagiarism." Most existing zero-watermarking technologies struggle to penetrate this semantic reconstruction and cannot identify the inheritance relationship between the rewritten text and the original. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned shortcomings in the prior art by providing a zero-watermark generation and detection method based on multi-dimensional text features. The aim is to construct a zero-watermark system with strong anti-attack capabilities and high detection accuracy by integrating features from three levels: vocabulary, sentence, and document, thereby solving the defects existing in the prior art.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A zero-watermark generation and detection method based on multi-dimensional text features includes the following steps: S1. Preprocess and semantically normalize the input text to obtain normalized text; S2. Extract multi-dimensional text features from the normalized text to construct a zero watermark; S3. Generate a zero-watermark identifier that is associated with and bound to multi-dimensional text features, and store the zero-watermark identifier; S4. During copyright verification, extract multi-dimensional text features from the text to be detected and calculate the similarity between the text and each dimension feature in the stored zero-watermark identifier. S5. Construct a joint feature vector from the similarity between features of each dimension, input it into the classification model for weighted fusion judgment, and output the copyright authentication result of the text to be detected.

[0009] Furthermore, step S1 includes the following sub-steps: S11. Perform structured processing on the input raw text, identify paragraph marks to segment it, and process each paragraph into sentences according to punctuation marks to obtain a sentence sequence; S12. Perform lexical filtering on the text after sentence segmentation, and use a preset stop word list to remove invalid words from the text; S13. Perform semantic normalization processing on the text after word filtering. Based on the synonym mapping relation library, use the longest matching strategy to map words with the same or similar meanings to unified standard words. S14. After the semantic normalization of synonyms, the text is reorganized according to the original paragraph and sentence order to generate a standardized text.

[0010] Furthermore, step S2 includes the following sub-steps: S21. Extracting lexical-level features: Calculate the weight value of each word in the normalized text using the term frequency-inverse document frequency algorithm, select the top K words with the highest weight values ​​as keywords to generate a keyword set, and construct a lexical-level zero-watermark feature vector composed of keywords and their weights. S22. Extract sentence-level features: First, calculate the importance features of the sentence in the statistical dimension, which includes at least sentence length, sentence position, and keyword density in the sentence; at the same time, use a pre-trained language model to calculate the semantic attention score of the sentence in the context; weight and fuse the importance features in the statistical dimension with the semantic attention score to obtain the comprehensive score of the sentence, select the top K sentences with the highest scores, convert them into a set of semantic vectors, and use them as sentence-level zero-watermark features; S23. Extract document-level features: Generate a text summary representing the core content of the input document using a generative large language model, and then encode the text summary using a semantic coding model to convert it into a document-level feature vector in a high-dimensional semantic vector space.

[0011] Furthermore, step S22 includes the following sub-steps: S221. Calculate the importance features of the statistical dimensions of sentence length, sentence position in paragraph, and keyword density in sentence; S222. Calculate the semantic attention score of the sentence in the context: Obtain the last three layers of attention matrices output by the BERT model, calculate the global mean of each layer of attention matrix, and sum the global mean of the last three layers according to the preset weight coefficients to obtain the semantic attention score of the sentence in the context. S223. Normalize and weight the statistical importance features of sentence length, sentence position in paragraph, keyword density in sentence, and semantic attention score to calculate the comprehensive score of each sentence. S224. Select the top K sentences with the highest overall scores, and use the pre-trained encoding model BERT to map the corresponding sentences into feature vectors, thereby forming sentence-level zero-watermark features.

[0012] Furthermore, in step S221, a Gaussian kernel function is used to calculate the statistical dimension importance feature of sentence length, which is expressed as: In the formula, Score the importance of sentence length; The length of the character string. For sentences, The average sentence length For adaptive standard deviation; The importance feature of a sentence's position within a paragraph is calculated using statistical dimensions, and is expressed as follows: In the formula, Score the importance of the sentence's position in the paragraph. For sentence position index, The total number of sentences in the paragraph; The importance feature of the statistical dimension of keyword density in a sentence is calculated and expressed as follows: In the formula, The importance score is given based on the density of keywords in the sentence. For the first i One keyword, For a set of keywords, This represents the total number of words after the sentence is segmented.

[0013] Furthermore, in S222, the semantic attention score of the sentence in the context is calculated, which is expressed as: In the formula, Score the semantic attention of the sentence in the context. For the first sentence One keyword, The global mean of the BERT attention matrix is ​​calculated. This represents the total number of layers in the BERT model.

[0014] Furthermore, step S3 includes the following sub-steps: S31. Randomly generate a globally unique document identifier ID for the input document; S32. Associate and bind the unique document identifier ID with the word-level, sentence-level, and document-level features to jointly generate a structured zero-watermark identifier; S33. Submit the zero-watermark identifier to a third-party certification center for storage and registration to complete copyright registration.

[0015] Furthermore, step S4 includes the following sub-steps: S41. Calculate the lexical cosine similarity between the stored zero-watermark lexical features and the lexical features of the text to be detected. If the lexical similarity is lower than the preset threshold, the text is directly determined to be dissimilar; otherwise, continue the subsequent calculation. S42. Calculate the sentence-level similarity between the stored zero-watermark sentence-level features and the sentence-level features of the text to be detected; specifically, calculate the cosine similarity between each sentence vector in the zero-watermark and all sentence vectors in the text to be detected, take the maximum matching value, and finally calculate the average of all maximum matching values ​​as the sentence-level similarity. S43. Calculate the document-level cosine similarity between the stored zero-watermark document-level features and the document-level features of the text to be detected.

[0016] Furthermore, step S5 includes the following sub-steps: S51. Construct a joint feature vector from word-level similarity, sentence-level similarity, and document-level similarity; S52. Optimize the training classification model; based on models with real binary classification labels. training sample set Optimize and train the classification model; among which, For the first The joint feature vector of each sample For the first The true binary classification labels in each sample; S53. The classification model calculates the weighted fusion score by linearly combining the weight parameters and the joint feature vector, and outputs the copyright authentication result of the text to be detected by comparing the score with the probability threshold.

[0017] The zero-watermark generation and detection method based on multi-dimensional text features provided by this invention has the following beneficial effects: 1) Multi-dimensional feature fusion improves the robustness and accuracy of detection: This invention abandons the single-dimensional feature extraction method and innovatively integrates vocabulary-level statistical features, sentence-level hybrid semantic features, and document-level summary semantic features. This multi-granular design effectively solves the problem that traditional methods cannot simultaneously address both "resistance to synonym substitution" and "resistance to subtle data tampering".

[0018] 2) Enhanced ability to resist structural attacks: By introducing a sentence selection mechanism based on BERT attention and statistical features, combined with a maximum matching strategy based on vector sets, this invention can effectively deal with structural attacks such as paragraph rearrangement, sentence disorder and local cutting, and solves the problem of zero watermark synchronization failure when the text structure changes.

[0019] 3) Effective defense against generative rewriting attacks: By using a large language model (Qwen) to generate document summaries and extract semantic features, and by leveraging the high semantic consistency of the summaries before and after rewriting, the problem that traditional zero-watermarking technology cannot effectively detect AI-powered plagiarism is solved, thus adapting to the copyright protection needs in the current context of generative artificial intelligence. Attached Figure Description

[0020] Figure 1 This is a flowchart of the zero-watermark generation and detection method based on multi-dimensional text features in this embodiment. Detailed Implementation

[0021] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0022] This embodiment presents a zero-watermark generation and detection method based on multi-dimensional text features. It mainly includes four core stages: text preprocessing, multi-dimensional feature extraction, zero-watermark registration and evidence storage, and infringement detection and judgment. This method improves the robustness and accuracy of text detection, enhances the ability to resist structural attacks, and achieves effective defense against generative rewriting attacks. (Reference) Figure 1 Specifically, it includes the following: S1. Preprocess and semantically normalize the input text to obtain normalized text, eliminating format differences and noise interference; this specifically includes the following steps: S11. Perform structured processing on the input raw text, identify line breaks or paragraph marks in the text to segment it, and use periods, question marks, exclamation marks or semicolons as delimiters to process sentences in each paragraph to obtain a sentence sequence. S12. Perform lexical filtering on the text after sentence segmentation, load a preset stop word list (such as Baidu stop word list) to perform part-of-speech tagging on the text, and remove invalid words such as auxiliary words, prepositions, conjunctions and special symbols that have low contribution to the semantics of the text, so as to reduce the impact of noise words on subsequent feature extraction. S13. Perform semantic normalization processing on the text after word filtering. Based on the synonym mapping relation library, use the longest matching strategy to map words with the same or similar meanings to unified standard words. In some embodiments, based on a pre-built thesaurus containing domain-specific terminology (such as the Harbin Institute of Technology Thesaurus), different words with the same or similar meanings are mapped to unified standard words; wherein, in the thesaurus matching process, the longest matching strategy is executed by using either the forward maximum matching algorithm or the reverse maximum matching algorithm, prioritizing the matching of longer words to avoid semantic ambiguity; thus, it can resist thesaurus substitution attacks.

[0023] S14. After semantic normalization of synonyms, the text is reorganized according to the original paragraph and sentence order to generate standardized text, which is used for subsequent extraction of multi-dimensional text features and construction of zero watermark.

[0024] S2. Extract multi-dimensional text features from normalized text to construct a zero watermark; this specifically includes the following steps: S21. Extracting Lexical-Level Features: The Term Frequency-Inverse Document Frequency (TF-IDF) algorithm is used to calculate the weight of each word in the normalized text. The top K words with the highest weights are selected as keywords, and a keyword set is generated. Finally, a lexical-level zero-watermark feature vector is constructed, consisting of the keywords and their weights. ; In some embodiments, for words in normalized text Its weight value is calculated as follows: In the formula, Words Frequency of appearance in the text; For vocabulary The number of times it appears in normalized text. To normalize the frequency of the k-th word in the text, This is an index variable used to iterate through all words in the normalized text. Inverse document frequency (IVF) is used to measure vocabulary. The importance of; in, Represented as: In the formula, N This indicates the total number of texts in the corpus. Indicates included words The number of texts.

[0025] Based on the calculated TF-IDF weight values, all words are sorted, and the top K words by weight value are selected as the keyword set W={ }, and construct a vocabulary-level zero-watermark feature vector: in, For the first One keyword, Keywords The corresponding TF-IDF weight values.

[0026] S22. Extracting Sentence-Level Features: Analyze sentences in the normalized text. First, calculate the importance features of sentences in statistical dimensions, which include at least sentence length, sentence position, and keyword density. Simultaneously, use a BERT pre-trained language model to calculate the semantic attention score of the sentence in its context. Weightedly fuse the statistical importance features with the semantic attention score to obtain a comprehensive sentence score. Select the top K sentences and convert them into a set of semantic vectors using the BERT pre-trained model, which will then serve as sentence-level zero-watermark features. This process includes the following steps: S221. Calculate the importance features of the statistical dimensions of sentence length, sentence position in paragraph, and keyword density in sentence; Importance features of sentence length statistical dimension; Based on the normalized text, identify the sentence sequence within it. M represents the total number of sentences; calculate the sum of each sentence. character length Based on the length distribution of all sentences, a Gaussian kernel function is used to calculate the importance score of sentence length, which is expressed as: In the formula, Score the importance of sentence length; The length of the character string. For sentences, The average sentence length For adaptive standard deviation; importance score of sentence length. It reflects the degree to which the sentence length deviates from the average; the closer the length is to the average, the higher the score.

[0027] The importance feature of a sentence's position within a paragraph, based on statistical dimensions, is expressed as follows: In the formula, Score the importance of the sentence's position in the paragraph. For sentence position index, The total number of sentences in the paragraph; The importance feature of the statistical dimension of keyword density in a sentence is expressed as follows: In the formula, The importance score is given based on the density of keywords in the sentence. For the first i One keyword, For a set of keywords, The total number of words after sentence segmentation; This indicates the number of keywords in a sentence; in this way, the higher the keyword density in a sentence, the higher the score.

[0028] S222. Calculate the semantic attention score of the sentence in the context; Obtain the last three attention matrices output by the BERT model, calculate the global mean of each attention matrix, and then sum the global means of the last three layers according to preset weight coefficients to obtain the semantic attention score of the sentence in the context, which is expressed as follows: In the formula, Score the semantic attention of the sentence in the context. For the first sentence One keyword, The global mean of the BERT attention matrix is ​​calculated. The total number of layers in the BERT model S223, Sentence-level watermark generation: Get the generated score Next, normalization is performed first, and then a comprehensive score is calculated for each sentence, with the weight coefficients for the four features set as follows: The score for sentence j is: In the formula, The overall score is calculated based on the total score.

[0029] S224. Based on the overall score Sort all sentences in descending order and select the top K sentences. Using the pre-trained encoding model BERT Each sentence in the document is mapped to a feature vector, forming sentence-level zero-watermark features. ;in, For the first Semantic vectors of key sentences.

[0030] S23. Extracting Document-Level Features: A generative large language model (qwen) is used to generate a text summary representing the core content of the input document. Then, a semantic encoding model (SentenceTransformer) is used to encode the text summary, converting it into a document-level feature vector in a high-dimensional semantic vector space. .

[0031] S3. Generate and store zero-watermark identifiers that are associated with and bound to multi-dimensional text features; this includes the following sub-steps: S31. Randomly generate a globally unique document identifier ID for the input document; S32. Associate and bind the unique document identifier (ID) with word-level, sentence-level, and document-level features to generate a structured zero-watermark identifier. ; S33. Submit the zero-watermark mark to a third-party certification center for storage and registration to complete copyright registration.

[0032] S4. During copyright verification, multi-dimensional text features are extracted from the text to be detected, and the similarity between these features and the features of each dimension in the stored zero-watermark identifier is calculated. This process includes the following steps: S41. Computationally stored zero-watermark vocabulary-level features Lexical features of the text to be detected Lexical cosine similarity between : In the formula, Represents the dot product of two vectors. The L2 norm of a vector; Set a fixed threshold ,like If the two texts are completely unrelated at the word level, the overall similarity is returned as 0, and the calculation of subsequent sentence-level similarity and document-level similarity is skipped. S42. Calculate the corresponding sentence-level features for the zero-watermark stored in the third-party authentication center. Sentence-level features of the text to be detected The maximum matching strategy is used to calculate the similarity, that is, for... any vector in Calculate its relationship with Calculate the cosine similarity of all vectors in the sentence and take the maximum value as the matching score for that vector; calculate the average of the matching scores of all K vectors to obtain the sentence-level similarity. ; S43. Calculate the document-level cosine similarity between the stored zero-watermark document-level features and the document-level features of the text to be detected; For zero-watermarks stored in a third-party certification center, calculate their corresponding document-level features. Document-level features of the text to be detected Cosine similarity between The calculation method is as follows, where Represents the dot product of two vectors. The L2 norm of a vector is used to calculate document-level similarity. ; S5. Construct a joint feature vector from the similarities between features of each dimension, and input it into the classification model for weighted fusion determination, outputting the copyright authentication result of the text to be detected. This specifically includes the following steps: S51. Construct a joint feature vector from word-level similarity, sentence-level similarity, and document-level similarity; In some embodiments, word-level similarity Sentence-level similarity Document-level similarity The similarity scores from these three dimensions are concatenated to form a joint feature vector. Learning joint eigenvectors using a linear regression model. With binary labels The mapping relationship between them.

[0033] The linear regression model is expressed as follows: In the formula, The results of the model classification, Let be the weight vector to be learned, and let its components correspond to the weights of word-level, sentence-level, and document-level features, respectively. For bias terms, This is the Sigmoid activation function.

[0034] S52. Optimize the training classification model; based on models with real binary classification labels. training sample set The optimal weight vector is obtained by minimizing the cross-entropy loss function using gradient descent. and bias terms ;in, For the first The joint feature vector of each sample For the first The true binary classification labels in each sample; S53. Extracting word-level similarity Sentence-level similarity Document-level similarity The classification model calculates a weighted score by linearly combining weight parameters and joint feature vectors. The copyright authentication result of the text to be detected is output by comparing the score with the probability threshold. If the score A score greater than 0.5 indicates that the two texts are similar; otherwise, they are not similar.

[0035] Although specific embodiments of the invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of this patent. Various modifications and variations that can be made by a person skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of this patent.

Claims

1. A method for zero-watermark generation and detection based on multi-dimensional text features, characterized in that, Includes the following steps: S1. Preprocess and semantically normalize the input text to obtain normalized text; S2. Extract multi-dimensional text features from the normalized text to construct a zero watermark; S3. Generate a zero-watermark identifier that is associated with and bound to multi-dimensional text features, and store the zero-watermark identifier; S4. During copyright verification, extract multi-dimensional text features from the text to be detected and calculate the similarity between the text and each dimension feature in the stored zero-watermark identifier. S5. Construct a joint feature vector from the similarity between features of each dimension, input it into the classification model for weighted fusion judgment, and output the copyright authentication result of the text to be detected.

2. The zero-watermark generation and detection method based on multi-dimensional text features according to claim 1, characterized in that, S1 includes the following steps: S11. Perform structured processing on the input raw text, identify paragraph marks to segment it, and process each paragraph into sentences according to punctuation marks to obtain a sentence sequence; S12. Perform lexical filtering on the text after sentence segmentation, and use a preset stop word list to remove invalid words from the text; S13. Perform semantic normalization processing on the text after word filtering. Based on the synonym mapping relation library, use the longest matching strategy to map words with the same or similar meanings to unified standard words. S14. After the semantic normalization of synonyms, the text is reorganized according to the original paragraph and sentence order to generate a standardized text.

3. The zero-watermark generation and detection method based on multi-dimensional text features according to claim 1, characterized in that, S2 includes the following steps: S21. Extracting lexical-level features: Calculate the weight value of each word in the normalized text using the term frequency-inverse document frequency algorithm, select the top K words with the highest weight values ​​as keywords to generate a keyword set, and construct a lexical-level zero-watermark feature vector composed of keywords and their weights. S22. Extract sentence-level features: First, calculate the importance features of the sentence in the statistical dimension, which includes at least sentence length, sentence position, and keyword density in the sentence; at the same time, use a pre-trained language model to calculate the semantic attention score of the sentence in the context. The importance features of the statistical dimension are weighted and fused with the semantic attention score to obtain the comprehensive score of the sentence. The top K sentences with the highest scores are selected, converted into a set of semantic vectors, and used as sentence-level zero-watermark features. S23. Extract document-level features: Generate a text summary representing the core content of the input document using a generative large language model, and then encode the text summary using a semantic coding model to convert it into a document-level feature vector in a high-dimensional semantic vector space.

4. The zero-watermark generation and detection method based on multi-dimensional text features according to claim 3, characterized in that, S22 includes the following sub-steps: S221. Calculate the importance features of the statistical dimensions of sentence length, sentence position in paragraph, and keyword density in sentence; S222. Calculate the semantic attention score of the sentence in the context: Obtain the last three layers of attention matrices output by the BERT model, calculate the global mean of each layer of attention matrix, and sum the global mean of the last three layers according to the preset weight coefficients to obtain the semantic attention score of the sentence in the context. S223. Normalize and weight the statistical importance features of sentence length, sentence position in paragraph, keyword density in sentence, and semantic attention score to calculate the comprehensive score of each sentence. S224. Select the top K sentences with the highest overall scores, and use the pre-trained encoding model BERT to map the corresponding sentences into feature vectors, thereby forming sentence-level zero-watermark features.

5. The zero-watermark generation and detection method based on multi-dimensional text features according to claim 4, characterized in that, In step S221, a Gaussian kernel function is used to calculate the statistical dimension importance feature of sentence length, which is expressed as: In the formula, Score the importance of sentence length; The length of the character string. For sentences, The average sentence length For adaptive standard deviation; The importance feature of a sentence's position within a paragraph is calculated using statistical dimensions, and is expressed as follows: In the formula, The sentence is scored based on its importance within the paragraph. For sentence position index, The total number of sentences in the paragraph; The importance feature of the statistical dimension of keyword density in a sentence is calculated and expressed as follows: In the formula, The importance score is assigned based on the density of keywords in the sentence. For the first i One keyword, For a set of keywords, This represents the total number of words after the sentence is segmented.

6. The zero-watermark generation and detection method based on multi-dimensional text features according to claim 4, characterized in that, In step S222, the semantic attention score of the sentence in the context is calculated, which is expressed as: In the formula, Score the semantic attention of the sentence in the context. For the first sentence One keyword, The global mean of the BERT attention matrix is ​​calculated. This represents the total number of layers in the BERT model.

7. The zero-watermark generation and detection method based on multi-dimensional text features according to claim 1, characterized in that, S3 includes the following steps: S31. Randomly generate a globally unique document identifier ID for the input document; S32. Associate and bind the unique document identifier ID with the word-level, sentence-level, and document-level features to jointly generate a structured zero-watermark identifier; S33. Submit the zero-watermark identifier to a third-party certification center for storage and registration to complete copyright registration.

8. The zero-watermark generation and detection method based on multi-dimensional text features according to claim 1, characterized in that, S4 includes the following steps: S41. Calculate the lexical cosine similarity between the stored zero-watermark lexical features and the lexical features of the text to be detected. If the lexical similarity is lower than the preset threshold, the text is directly determined to be dissimilar; otherwise, continue the subsequent calculation. S42. Calculate the sentence-level similarity between the stored zero-watermark sentence-level features and the sentence-level features of the text to be detected; Specifically, the cosine similarity between each sentence vector in the zero watermark and all sentence vectors in the text to be detected is calculated, and the maximum matching value is taken. Finally, the average of all the maximum matching values ​​is calculated as the sentence-level similarity. S43. Calculate the document-level cosine similarity between the stored zero-watermark document-level features and the document-level features of the text to be detected.

9. The zero-watermark generation and detection method based on multi-dimensional text features according to claim 1, characterized in that, S5 includes the following steps: S51. Construct a joint feature vector from word-level similarity, sentence-level similarity, and document-level similarity; S52. Optimize the training classification model; based on models with real binary classification labels. training sample set Optimize and train the classification model; among which, For the first The joint feature vector of each sample For the first The true binary classification labels in each sample; S53. The classification model calculates the weighted fusion score by linearly combining the weight parameters and the joint feature vector, and outputs the copyright authentication result of the text to be detected by comparing the score with the probability threshold.

Images