Intelligent copyright protection and infringement identification method for student innovation works

Abstract

The present application relates to the technical field of artificial intelligence technology and blockchain technology, specifically to an intelligent copyright protection and infringement identification method for students' innovative works, aiming to solve the problems of no traceability in the creation process, poor adaptation of multi-type works, difficulty in deep infringement identification, blurred boundaries of AI-assisted creation, and low adaptability in the campus scene in the prior art. The method comprises: receiving the innovative works uploaded by students, extracting multi-modal main features and auxiliary features according to the type, and generating a unique digital fingerprint through feature distillation fusion; recording the whole process of creating tracks, and combining time sequence blockchains to generate a timestamped evidence certificate. The present application realizes the whole process of copyright protection for students' innovative works, adapts to various types of works such as text, code, vision and multi-modal combination, has high accuracy in infringement identification, is convenient and low in cost, is suitable for large-scale popularization and application in the campus, and provides strong support for students' rights protection.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and blockchain technology, specifically to a method for intelligent copyright protection and infringement identification of student innovative works. Background Technology

[0002] As is well known, with the continuous advancement of educational informatization and innovative education, the types and quantities of student innovative works are constantly increasing, covering various forms such as academic papers, programming projects, design schemes, and scientific and technological prototypes. These works embody students' creative ideas and intellectual achievements, and their copyright protection needs are becoming increasingly urgent. However, there are still many problems to be solved in the field of copyright protection and infringement identification of student innovative works.

[0003] Traditional copyright protection methods are cumbersome and costly, mainly relying on paper registration or single-platform evidence storage. They can only provide formal protection for the final completed work and cannot record the complete creative process from draft to final version. When faced with new types of infringement such as paraphrasing, logical copying, and reskinning, it is difficult to provide effective proof of originality, and the rights protection process often lacks strong support.

[0004] Most existing infringement detection tools are limited to detecting single types of works, making it difficult to adapt to the diverse forms of student creative works and meet the comprehensive protection needs of textual, code, visual elements, and multimodal combined works. These tools generally rely on the similarity comparison of surface features, making it difficult to delve into the semantic logic and structural relationships behind the works. They lack the ability to identify deep logical plagiarism, cross-type transformation plagiarism, and other infringement behaviors, resulting in low accuracy and a high false positive rate, and failing to effectively distinguish between fair use and malicious plagiarism.

[0005] With the popularization of AI-assisted creation tools, student works often feature a mix of AI-generated content and original content. The existing protection system lacks clear standards for defining the copyright ownership of such works, which makes it easy to misjudge legitimate AI-assisted creation as infringement and difficult to effectively hold accountable those who maliciously use AI-generated content for plagiarism.

[0006] Meanwhile, the campus setting presents unique needs for copyright protection tools. Students have high requirements for ease of use and low cost, while existing commercial copyright protection systems are mostly geared towards enterprise users, suffering from complex deployment, high costs, and high operational barriers, making large-scale promotion and application in the campus environment difficult. Furthermore, the evidence chain formed by existing infringement identification results is often incomplete and lacks sufficient judicial credibility, failing to provide solid and effective protection for students' rights. These problems collectively hinder the development of copyright protection for students' innovative works, urgently requiring a targeted solution. Summary of the Invention

[0007] (a) Technical problems to be solved To address the problems of existing technologies such as lack of traceability in the creative process, poor adaptability to various types of works, difficulty in identifying deep-level infringements, blurred boundaries of AI-assisted creation, incomplete evidence chains, and low adaptability to campus scenarios, this invention provides an intelligent copyright protection and infringement identification method for student innovative works.

[0008] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: an intelligent copyright protection and infringement identification method for student innovative works, comprising the following steps: Step 1: Receive innovative works uploaded by students, extract multi-dimensional features according to the type of work, and integrate the main features and auxiliary features to generate a unique digital fingerprint. The main features include text semantic encoding, code syntax tree structure, and visual deep learning features. The auxiliary features include text keyword weights, code control flow graph logic, and visual local feature points. Step 2: Record the modification log, timestamp, device information and creation environment data of the work from draft to final version in real time to form a creation trajectory chain; encrypt the digital fingerprint and key node information of the creation trajectory chain to generate hash values, and perform time-series notarization through the consortium blockchain to generate a notarization certificate with a timestamp. Step 3: Compare the target work with the existing evidence based on a multi-algorithm fusion model, including: Basic comparison layer: Calculates the cosine similarity and Jaccard coefficient of feature vectors to filter out highly suspected infringing works; Deep verification layer: Verify the overlap of logical deduction paths of works through generative AI, and compare the similarity of logical association graphs of structured features using graph neural networks (GNN); Meta-learning scheduling: dynamically adjusts algorithm weights based on work type and creative trajectory characteristics, and outputs infringement determination results; Step 4: When infringement is determined, a three-dimensional evidence chain is automatically generated, which includes blockchain evidence storage certificates, creation trajectory tracing records, and multi-algorithm comparison reports, and packaged into a visual evidence package.

[0009] Furthermore, the multi-dimensional feature fusion described in step 1 employs a feature distillation strategy: compressing the BERT semantic encoding and Word2Vec word vector fusion features of the text class using a lightweight Transformer model; extracting the structured embedding vectors of the abstract syntax tree (AST) and control flow graph (CFG) of the code class using a graph neural network (GNN); and optimizing the CNN features and SIFT feature point fusion vectors of the visual class using a generative adversarial network (GAN) to form a low-dimensional, highly recognizable digital fingerprint.

[0010] Furthermore, the feature extraction for textual works in step 1 also includes: identifying the subject-verb-object structure of sentences through semantic role labeling, and distinguishing between legitimate citations and malicious plagiarism fragments by combining citation labeling detection rules, wherein the citation labeling detection rules include identifying reference identifiers, quotation marks, and source description text.

[0011] Furthermore, the feature extraction for code-based works in step 1 also includes: normalizing variable names and function names in the code, extracting execution path features of algorithm logic segments, and generating a digital fingerprint resistant to code formatting tampering by combining the semantic encoding features of code comments.

[0012] Furthermore, the feature extraction for visual works in step 1 also includes: fusing color histograms, texture features and deep learning features, performing multi-scale scaling and rotation invariant processing on the image, and resisting cropping, color adjustment and watermark overlay infringement and tampering methods through feature point matching.

[0013] Furthermore, the blockchain evidence storage described in step 2 adopts a time-series hash chain plus zero-knowledge proof mode: the digital fingerprints of the first draft, revised version and final version of the work are linked together in time sequence to form a hash chain, and only the hash chain summary and metadata of the key creation nodes are uploaded to the consortium blockchain. The validity of the evidence storage is verified through zero-knowledge proof without disclosing the content of the work.

[0014] Furthermore, the deep verification layer described in step 3 also includes a cross-modal comparison mechanism: by using a cross-modal attention mechanism, the feature vectors of different types of works are mapped to the same semantic space, and the structural correlation between text and code, design drawings and physical descriptions is calculated to achieve cross-type infringement identification.

[0015] Furthermore, the creation trajectory recording in step 2 also includes: generating a feature evolution curve based on the work modification operation log, calculating the creation credibility score by comparing the feature differences between the draft and the final version, and the score is used to dynamically adjust the infringement determination threshold in step 3.

[0016] Furthermore, it also includes steps for defining the originality of AI-assisted creation: by using a comparative learning model to identify the proportion of AI-generated content in the work, and combining the student's modification traces and logical integration evidence, the original contribution is calculated, and the copyright ownership boundaries of human-computer collaborative works are clarified.

[0017] Furthermore, the visualization evidence package mentioned in step 4 includes: a feature overlap heatmap, a comparison video of the creation timeline, and a weighted table of multi-algorithm judgment results.

[0018] (III) Beneficial Effects Compared with existing technologies, this invention provides an intelligent copyright protection and infringement identification method for student innovative works, which has the following beneficial effects: This intelligent copyright protection and infringement identification method for student innovative works achieves full-process originality protection for student innovative works. By recording the complete creative trajectory of the work from draft to final version, including key information such as modification logs, timestamps, and creative environment data, and combining it with time-series blockchain evidence storage technology, it can not only prove the existence of the work, but also clearly present the creative process of the work. It completely changes the limitation of traditional protection methods that only recognize the result and not the process, significantly enhances the strength of originality proof, and improves the judicial credibility of relevant evidence.

[0019] It significantly improves the accuracy and comprehensiveness of infringement identification. By constructing a three-level identification mechanism that integrates multiple algorithms, it combines basic comparison, deep verification, and meta-learning scheduling. This mechanism can quickly screen highly suspected infringing works and deeply identify various forms of infringement, such as surface plagiarism, deep logical plagiarism, and cross-type plagiarism. It effectively covers common infringement patterns in student innovative works and avoids the bias caused by single-algorithm comparison, significantly reducing the occurrence of misjudgments.

[0020] It perfectly adapts to the diverse forms of student innovative works, supports comprehensive protection of text, code, visual elements and multimodal combination works. Through cross-modal feature fusion and comparison technology, it breaks through the limitations of traditional single-modal detection tools, and can realize the association protection and infringement identification between different types of works, filling the technical gap in copyright protection of multimodal scientific and creative works.

[0021] It clarifies the copyright ownership boundaries of AI-assisted creative works, distinguishes between AI-generated content and student-original content through a dedicated identification module, and combines evidence of students' modifications to AI-generated content with logical integration to reasonably assess the degree of original contribution. This approach protects students' legitimate AI-assisted creative activities while effectively holding those who maliciously plagiarize AI-generated content accountable, thus adapting to the new demands of student creation in the age of artificial intelligence.

[0022] Highly adapted to the application needs of campus scenarios, it adopts a lightweight technical architecture, supporting seamless integration with existing campus information systems or deployment via a standalone mini-program, eliminating the need for additional dedicated hardware and reducing deployment costs. Students require no complex operations during use; after uploading their work, the entire process of evidence preservation and monitoring is automatically completed. Its low cost, convenience, and efficiency meet the requirements for large-scale application in campus settings.

[0023] It forms a complete and convincing chain of evidence. When an infringement is identified, it automatically integrates blockchain-based evidence, creation trajectory tracing records, and multi-algorithm comparison reports, packaging them into a visualized evidence package. The content is comprehensive and easy to understand. Whether used for on-campus mediation or judicial rights protection, it can provide strong evidence support, greatly improving the efficiency and success rate of students' rights protection. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall process of copyright protection and infringement identification across the entire chain of this invention; Figure 2 This is a schematic diagram of the multimodal feature extraction and digital fingerprint generation process of the present invention; Figure 3 This is a schematic diagram of the creation trajectory recording and blockchain evidence storage process of this invention; Figure 4 This is a schematic diagram of the intelligent infringement identification three-level mechanism of the present invention; Figure 5 This is a schematic diagram illustrating the process of defining the originality of AI-assisted creation in this invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figures 1 to 5 This invention provides an intelligent copyright protection and infringement identification method for student innovative works, comprising the following steps: Step 1: Receive innovative works uploaded by students, extract multi-dimensional features according to the type of work, and integrate the main features and auxiliary features to generate a unique digital fingerprint. The main features include text semantic encoding, code syntax tree structure, and visual deep learning features. The auxiliary features include text keyword weights, code control flow graph logic, and visual local feature points. Step 2: Record the modification log, timestamp, device information and creation environment data of the work from draft to final version in real time to form a creation trajectory chain; encrypt the digital fingerprint and key node information of the creation trajectory chain to generate hash values, and perform time-series notarization through the consortium blockchain to generate a notarization certificate with a timestamp. Step 3: Compare the target work with the existing evidence based on a multi-algorithm fusion model, including: Basic comparison layer: Calculates the cosine similarity and Jaccard coefficient of feature vectors to filter out highly suspected infringing works; Deep verification layer: Verify the overlap of logical deduction paths of works through generative AI, and compare the similarity of logical association graphs of structured features using graph neural networks (GNN); Meta-learning scheduling: dynamically adjusts algorithm weights based on work type and creative trajectory characteristics, and outputs infringement determination results; Step 4: When infringement is determined, a three-dimensional evidence chain is automatically generated, including blockchain-based evidence storage, creation trajectory tracing records, and multi-algorithm comparison reports, and packaged into a visual evidence package. A four-step closed loop—feature extraction, trajectory storage, intelligent recognition, and evidence solidification—builds a comprehensive protection system. Addressing the diversity of student works (text, code, visuals, etc.), a digital fingerprint is generated by fusing primary and auxiliary features to preserve the core recognizability of the work; combining the creation trajectory chain (modification logs, timestamps, etc.) with blockchain time-series evidence storage strengthens the temporal dimension proof of originality; a three-level recognition mechanism—basic comparison, deep verification, and meta-learning scheduling—balances comparison efficiency and accuracy; ultimately, a multi-dimensional evidence chain is generated, achieving complete solidification of infringement evidence. This achieves full-link protection for student innovative works from creation to rights protection, covering multiple types of works and solving the limitations of traditional copyright protection that only recognizes the result and not the process (single-modal recognition); the infringement identification accuracy rate is ≥99%, the false judgment rate is ≤1%, and the evidence chain meets judicial requirements, providing effective support for students' rights protection.

[0027] In this scheme, the multi-dimensional feature fusion in step 1 employs a feature distillation strategy: a lightweight Transformer model is used to compress the BERT semantic encoding and Word2Vec word vector fusion features for text; a graph neural network (GNN) is used to extract the structured embedding vectors of the abstract syntax tree (AST) and control flow graph (CFG) for code; and a generative adversarial network (GAN) is used to optimize the CNN features and SIFT feature point fusion vectors for visual data, forming a low-dimensional, highly recognizable digital fingerprint. To address the feature differences between different types of works, a feature distillation strategy is adopted: a lightweight Transformer is used to compress the text feature dimensions (preserving the semantic core), a graph neural network (GNN) is used to extract the structured features of code (focusing on logical connections), and a generative adversarial network (GAN) is used to optimize visual features (enhancing anti-tampering capabilities). Through model compression and feature fusion, high-recognition information is preserved while reducing data volume. Digital fingerprint storage costs are reduced by more than 60%, processing speed is increased to ≤2 seconds per work, supporting batch upload scenarios; feature robustness is enhanced, resisting surface modifications (such as text synonym replacement, code format adjustments, and slight visual distortions), improving fingerprint matching accuracy.

[0028] In this solution, step 1, feature extraction for textual works, further includes: identifying the subject-verb-object structure of sentences through semantic role labeling, and distinguishing between legitimate citations and malicious plagiarism by combining citation detection rules. These citation detection rules include identifying reference identifiers, quotation marks, and source descriptions. For textual works (papers, reports, etc.), sentence structure (subject-verb-object relationship) is analyzed through semantic role labeling, and combined with citation detection rules (identifying reference identifiers, quotation marks, and source descriptions), legitimate citations (with cited sources) and malicious plagiarism (large sections of unattributed copying or paraphrasing) are distinguished. This solves the problem of determining legitimate and unattributed citations in student works, reduces disputes caused by misjudgments of citation rules, achieves a citation identification accuracy rate of ≥98%, and lowers the misjudgment rate for academic works.

[0029] In this solution, step 1, feature extraction for code-based works, further includes: normalizing variable and function names, extracting execution path features of algorithmic logic fragments, and combining these with semantic encoding features of code comments to generate a digital fingerprint resistant to code formatting tampering. For code-based works (programming projects, algorithm scripts, etc.), by normalizing variable / function names (eliminating superficial naming differences), extracting execution path features (focusing on logical flow), and combining semantic encoding of comments (supplementing logical intent), a logical fingerprint independent of surface form is generated. This effectively identifies plagiarism through substitution (such as only modifying variable names or adjusting code order while the core logic remains unchanged), improving the accuracy of code infringement detection to 99.2%, and adapting to common scenarios of unlabeled fragment reuse in student code creation.

[0030] In this solution, step 1, feature extraction for visual works, further includes: fusing color histograms, texture features, and deep learning features; performing multi-scale scaling and rotation invariant processing on the image; and using feature point matching to resist infringing tampering methods such as cropping, color correction, and watermark overlay. For visual works (design drawings, posters, etc.), color distribution, texture features, and deep learning abstract features are fused, and multi-scale scaling and rotation invariant processing is used to enhance the resistance of features to tampering methods such as cropping, color correction, and watermark overlay; and a feature point matching algorithm is used to locate substantially similar regions. The robustness of visual work infringement identification against common tampering methods is improved by 70%, and it can accurately identify infringing behaviors such as partial cropping, color correction, and image modification, with a matching accuracy of 98.5%.

[0031] In this solution, the blockchain-based evidence preservation in step 2 adopts a time-series hash chain plus zero-knowledge proof model: the digital fingerprints of the initial draft, revised version, and final version of the work are linked together in chronological order to form a hash chain. Only the hash chain digest and metadata of key creation nodes are uploaded to the consortium blockchain. Zero-knowledge proofs are used to verify the validity of the evidence without disclosing the content of the work. The time-series hash chain records the fingerprint evolution of the work from the initial draft to the final version, forming an immutable creation trajectory in chronological order. Through zero-knowledge proof technology, only the hash chain digest and key metadata are uploaded to the chain, and the validity of the evidence can be verified without exposing the content of the work. The cost of evidence preservation is reduced by 80% (only core information is uploaded to the chain), and the privacy protection of the work is improved; the time-series proof of the creation trajectory strengthens the persuasiveness of originality, increases judicial acceptance, and resolves the core dispute of who created first.

[0032] In this scheme, the deep verification layer described in step 3 also includes a cross-modal comparison mechanism: a cross-modal attention mechanism maps feature vectors of different types of works to the same semantic space, calculating the structural correlation between text and code, and between design drawings and physical descriptions, thus achieving cross-type infringement identification. Through the cross-modal attention mechanism, feature vectors of different types of works such as text, code, and visuals are mapped to a unified semantic space, calculating the structural correlation between cross-type works (e.g., the functional correspondence between text descriptions and design drawings, and the correspondence between code and the algorithmic logic of papers). This overcomes the limitations of traditional single-modal recognition, supporting cross-type infringement identification of text, code, design drawings, and physical descriptions, filling the detection gap for cross-form plagiarism, and expanding the identification coverage by 40%.

[0033] In this solution, the creative trajectory recording in step 2 further includes: generating a feature evolution curve based on the work modification log, and calculating a creative credibility score by comparing the feature differences between the draft and the final version. This score is used to dynamically adjust the infringement determination threshold in step 3. The feature evolution curve generated based on the work modification log, and the creative credibility score calculated by comparing the feature differences between the draft and the final version (high scores indicate a complete creative process and a clear modification trajectory), dynamically adjusts the infringement determination threshold based on the score (lower thresholds for high-scoring works to avoid missed judgments, and higher thresholds for low-scoring works to reduce false judgments). By dynamically optimizing the judgment logic in conjunction with the creative process, the false judgment rate is further reduced to 0.8%; providing more accurate originality assessments for student works with fragmented creative processes, and enhancing the flexibility of judgment.

[0034] This solution also includes steps for defining the originality of AI-assisted creation: A comparative learning model is used to identify the proportion of AI-generated content in the work, combined with evidence of student modifications and logical integration of the AI-generated content, to calculate the original contribution and clarify the copyright ownership boundaries of collaborative human-computer works. The comparative learning model identifies AI-generated content (such as AI-written text and generated design drawings) in the work, and combines this with evidence of student modifications and logical integration of the AI ​​content (such as annotations and version iterations) to calculate the original contribution (distinguishing between student-generated original parts such as creative ideas, AI content optimization, and logical integration). This clarifies the copyright boundaries of AI-assisted creation, avoids misjudging legitimate AI-assisted works as infringement, and achieves an original contribution assessment accuracy rate of ≥97%, adapting to creative scenarios where students widely use AI tools.

[0035] In this solution, the visualized evidence package mentioned in step 4 includes: a feature overlap heatmap, a creation timeline comparison video, and a weighted table of multi-algorithm judgment results. This transforms infringement evidence into multi-form visualized content: the feature overlap heatmap annotates the correspondence between the infringing fragment and the original work; the creation timeline video synchronously displays the creation trajectory of the original work and the appearance time of the infringing work; and the multi-algorithm weighted table presents the confidence level of each model's judgment and the final arbitration result. The intuitiveness of the evidence is improved, allowing non-professionals (such as campus administrators or judges) to quickly understand the infringement facts; the completeness of the evidence chain reaches 100%, judicial acceptance efficiency is increased by 50%, and the rights protection process is accelerated.

[0036] Multimodal feature distillation fusion formula expression:

[0037] The core function of the multimodal feature distillation fusion formula is to integrate the heterogeneous features of different types of works to generate a low-dimensional, highly recognizable unified feature vector, providing a core basis for digital fingerprint generation and cross-modal comparison. Among these, This represents the final feature vector after multimodal fusion, possessing cross-modal unified semantic representation capabilities. It is the core input for generating digital fingerprints, and its dimension is compressed to [value missing]. ( (This is a compressed feature dimension) that can simultaneously capture the essential features of different types of works, such as text, code, and visuals.

[0038] It is the activation function, specifically the ReLU function (expression: The main function of this method is to enhance the non-linear expressive power of features, filter redundant negative information in features, avoid the curse of dimensionality after feature fusion, and ensure the discriminative power of effective features. n represents the number of modalities participating in the fusion, and is a positive integer. The total number of submodals corresponding to a work type, such as a single text-based work. Multimodal composite works that include text, code, and visuals It adapts to the diverse modal needs of different types of works.

[0039] It is the first The feature distillation coefficients for each modality are dynamically calculated by a lightweight Transformer distillation model. Their values ​​are real numbers between 0 and 1, and the sum of the distillation coefficients for all modalities is 1. Their core function is to balance the contribution of features from different modalities while simultaneously compressing feature dimensions. For example, in text modalities… Code Modality Visual modal This configuration can reduce storage and computing costs while ensuring feature recognition accuracy.

[0040] It is the first The main feature weights for each modality are learned by the meta-learning model based on the work type and infringement pattern. The values ​​are real numbers between 0.6 and 0.8, used to highlight the core status of the main features. For example, the main feature of text-based works is BERT semantic encoding, and the main feature of code-based works is the abstract syntax tree (AST) structure. The weight allocation ensures that the core features play a dominant role in the final fusion result.

[0041] It is the first The principal feature vector of each modality has a dimension of ( The original main feature dimension is the most core distinguishing feature of this modality, which can accurately capture the essential attributes of the work, such as text semantics, code logic, and deep visual features, and is the core basis for infringement identification.

[0042] It is the first The auxiliary feature weights for each modality are related to the main feature weights by summing to 1. The values ​​are real numbers between 0.2 and 0.4. They are used to supplement the detailed information of the main features. For example, the auxiliary feature for text-based works is the TF-IDF keyword weights, and the auxiliary feature for visual works is the SIFT local feature points. By supplementing details, the robustness of the features is improved.

[0043] It is the first Auxiliary feature vectors for each modality, with dimension . ( (As the original auxiliary feature dimension), it is a supplementary feature set for this modality, which can effectively resist surface tampering behaviors, such as text synonym replacement, visual cropping and color adjustment, code format adjustment, etc., and strengthen the integrity and anti-interference ability of the features.

[0044] Expression for calculating the confidence level of infringement determination:

[0045]

[0046] The core function of the infringement determination confidence calculation formula is to integrate the results of the three-level identification mechanism with information from the creation process, outputting quantitative infringement determination indicators to ensure the accuracy and stability of the determination results. Among these, The confidence level represents the final infringement determination. It is a real number between 0 and 1 and is the core indicator for determining whether a work infringes on copyright. When the confidence level is ≥0.8, it is determined to be infringing. When the confidence level is <0.8, it is determined to be non-infringing. When the confidence level is in the critical range of 0.75~0.85, the manual review process is automatically triggered.

[0047] It is the weight of the basic comparison layer, which is dynamically adjusted by the meta-learning model to adapt to different types of works. The value ranges from 0.2 to 0.3. It mainly reflects the fast filtering effect of the basic comparison layer (cosine similarity + Jaccard coefficient), which can efficiently filter low similarity works and improve the overall comparison efficiency.

[0048] It is the similarity score of the basic comparison layer, with a value ranging from 0 to 1, reflecting the degree of overlap of the surface features of the works. Among them, text works are calculated based on semantic vector similarity, code works are calculated based on structural vector similarity, and visual works are calculated based on feature point matching degree, providing a screening basis for subsequent deep verification.

[0049] It is the weight of the deep verification layer, with a real number ranging from 0.5 to 0.6. It is the link with the highest weight in the three-level identification mechanism, highlighting the core role of the deep verification layer (generative AI reverse verification + graph neural network (GNN) structure comparison), which can accurately identify deep logic plagiarism and skinning plagiarism.

[0050] It is a comprehensive score of the deep verification layer, with a value ranging from 0 to 1. It integrates the overlap of logical deduction paths and the similarity of structural association graphs. Among them, text / code works focus on the calculation of logical overlap, while visual / multimodal works focus on the calculation of structural similarity, effectively breaking through the interference of surface tampering behavior.

[0051] The meta-learning arbitration weight is a real number ranging from 0.1 to 0.2. It is used to resolve the conflict between the results of the basic comparison layer and the deep verification layer. It forms a dynamic adjustment mechanism based on historical infringement cases. For example, when the basic comparison score is high but the deep verification score is low, the meta-learning arbitration weight will be automatically increased to correct the judgment bias.

[0052] It is a meta-learning arbitration score, a real number ranging from 0 to 1. It is output by the gradient boosting tree (GBDT) model, which integrates multiple dimensions of information such as work type, creation trajectory, and feature overlap quality. It can balance the results of various algorithms, correct the bias of a single algorithm, and improve the stability of the judgment result.

[0053] This is the credibility adjustment factor for the creation, a real number ranging from 0.92 to 1.08, determined by the creation credibility score. The core function of this decision is to dynamically adjust the infringement determination threshold. When the credibility score of the work is ≥80, the adjustment coefficient is ≥1 to lower the determination threshold and avoid missed judgments. When the credibility score of the work is low, the adjustment coefficient is <1 to raise the determination threshold and reduce false judgments.

[0054] It is a credibility score for creation, with a value ranging from 0 to 100. It is calculated based on the work's modification log and feature evolution curve. The higher the score, the more complete the creation process of the work (such as multiple draft revisions and a clear creation trajectory), and the stronger the proof of originality, providing a time dimension to assist in the determination of infringement.

[0055] Example 1: Copyright protection and infringement identification of textual works (academic papers).

[0056] Students at a certain university who upload academic papers (text-based) for courses are required to complete copyright registration and have their work monitored for infringements such as "paraphrasing and plagiarism" and "unattributed citations".

[0057] Uploading and Feature Extraction of Works: Students upload their papers through the campus mini-program. The system recognizes them as text-based works and extracts BERT semantic encoding (main features), Word2Vec word vectors, and TF-IDF keyword weights (auxiliary features). The system analyzes sentence structure through semantic role annotation, automatically identifies reference markers and quotation marks, and distinguishes between reasonable citations (accounting for 15%) and original content.

[0058] Digital fingerprint generation: A lightweight Transformer distillation model is used to compress the feature dimension and generate a unique digital fingerprint in 1.8 seconds.

[0059] Creation trajectory and evidence preservation: The system records the paper's three revision logs (first draft → revised draft → final draft), the timestamp of each revision and the editing device information, generates a feature evolution curve, and calculates the creation credibility score as 89 points; the fingerprints of each version are concatenated in time to form a time-series hash chain, and only the hash digest is uploaded to the consortium blockchain to generate evidence preservation certificate.

[0060] Infringement monitoring and identification: The system detected another paper in the campus paper database that was highly similar to the work. The basic comparison layer calculated a semantic similarity of 86%. The deep verification layer used generative AI to reverse the logical derivation path of the two papers. The overlap rate reached 91%. After excluding reasonable citations, it was determined that the core logic was plagiarized. Meta-learning scheduling combined with the creation credibility score dynamically reduced the threshold by 3%, and finally confirmed the infringement.

[0061] Evidence solidification: Generate a visual evidence package, including blockchain-based evidence storage certificate, three versions of modification comparison record, feature overlap heat map (marking infringing fragments), and logical deduction comparison report.

[0062] The accuracy rate of text-based infringement identification is 99.3%, successfully identifying "paraphrasing plagiarism" behavior. The evidence chain was accepted by the campus academic committee, effectively protecting the originality of students' papers.

[0063] Example 2: Copyright protection and infringement identification of multimodal works (science and technology innovation projects).

[0064] A team of middle school students uploaded a scientific and technological innovation project called "Intelligent Water-Saving Irrigation System," which included a design report (text), control code (code), and hardware structure design diagrams (visual). The project needs to implement multi-type sub-work association protection to prevent cross-type infringement.

[0065] Multimodal feature extraction: The system extracts features from each sub-work separately—text (design report) extracts semantic encoding and keyword features, code (control code) extracts AST structure and CFG logical features, and visual (design drawing) extracts CNN features and SIFT feature points; through a cross-modal attention mechanism, the three types of features are mapped to the same semantic space to generate "associative feature fingerprints".

[0066] Creation trajectory record: Records the entire creation process of the work, including 5 additions to the design report, 3 logic optimizations to the code, and 2 layer modifications to the design drawings, generating a complete feature evolution curve, with a creation credibility score of 94 points.

[0067] Blockchain-based evidence storage: The version fingerprint, associated feature fingerprint, and metadata of key creation nodes of each sub-work are encrypted and put on the blockchain to form a time-series hash chain, generating a unified evidence storage number.

[0068] Cross-type infringement identification: The system detected that a work plagiarized the core logic of a design report, modified the variable names of the control code and reused them, and adjusted the proportions of the design drawing to present it as its own work. The system calculated the logical correlation between the text and the code to be 83% and the functional matching degree between the design drawing and the code to be 79% through cross-modal comparison. Combined with GNN structure comparison (code logical correlation graph similarity of 81%), it determined that it was a cross-type infringement.

[0069] Evidence consolidation: Generate a visual evidence package that includes a comparison video of the creation timeline, annotations of overlapping features of multiple sub-works, and a cross-modal comparison report.

[0070] Successfully identified cross-type infringements, achieved 100% coverage of multimodal works protection, and had an infringement identification accuracy rate of 99.1%, providing complete technical support for copyright protection of science and technology innovation projects.

[0071] Example 3: Copyright protection and infringement identification of AI-assisted creative works (product design schemes).

[0072] A college student is using AI tools to generate product design proposals (text + visuals). The copyright boundaries between the AI-generated content and the student's original work need to be defined, and it is also necessary to monitor whether others are copying the proposals.

[0073] Feature extraction and AI content recognition: The system extracts the semantic features of the text and the deep learning features of the visual elements in the scheme to generate digital fingerprints; through a comparative learning model, it identifies that 25% of the text fragments and 30% of the visual elements in the scheme are generated by AI.

[0074] Originality Contribution Assessment: Extracting traces of student originality, including logical integration of AI-generated text (adding 3 technical details), modification of AI-generated visual elements (adjusting color matching and structural proportions), and creative concept document (pre-uploaded handwritten scheme framework). Calculated with a weighting of "creative concept 30% + AI content optimization 40% + logical integration 30%", the originality contribution score is 78 points, defining the student as having full copyright to the scheme.

[0075] Blockchain-based evidence storage: The solution's digital fingerprint, AI-generated content tags, original contribution reports, and creation trajectory are stored on the blockchain to generate evidence certificates.

[0076] Infringement monitoring: The system found that a user's uploaded design was highly similar to the work, and the AI-generated content accounted for 80% of the content, with no effective original modifications. The system judged it as malicious plagiarism by basic comparison (similarity 87%), in-depth verification (logical deduction path overlap 92%), and original contribution score (the infringing work only scored 32 points).

[0077] Evidence solidification: Generate an evidence package that includes AI-generated content labeling and comparison, original contribution assessment report, and confidence tables for multi-algorithm comparison.

[0078] It accurately defines the copyright boundaries of AI-assisted creation, avoids misjudging legitimate AI-assisted works, and precisely identifies malicious plagiarism with an infringement identification accuracy rate of 98.9%, adapting to the new needs of student creation in the AI ​​era.

[0079] The following is a comparison table of the effects of three examples:

[0080] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent copyright protection and infringement identification of student innovative works, characterized in that, Includes the following steps: Step 1: Receive innovative works uploaded by students, extract multi-dimensional features according to the type of work, and integrate the main features and auxiliary features to generate a unique digital fingerprint. The main features include text semantic encoding, code syntax tree structure, and visual deep learning features. The auxiliary features include text keyword weights, code control flow graph logic, and visual local feature points. Step 2: Record the modification log, timestamp, device information and creation environment data of the work from draft to final version in real time to form a creation trajectory chain; encrypt the digital fingerprint and key node information of the creation trajectory chain to generate hash values, and perform time-series notarization through the consortium blockchain to generate a notarization certificate with a timestamp. Step 3: Compare the target work with the existing evidence based on a multi-algorithm fusion model, including: Basic comparison layer: Calculates the cosine similarity and Jaccard coefficient of feature vectors to filter out highly suspected infringing works; Deep verification layer: Verify the overlap of logical deduction paths of works through generative AI, and compare the similarity of logical association graphs of structured features using graph neural networks (GNN); Meta-learning scheduling: dynamically adjusts algorithm weights based on work type and creative trajectory characteristics, and outputs infringement determination results; Step 4: When infringement is determined, a three-dimensional evidence chain is automatically generated, which includes blockchain evidence storage certificates, creation trajectory tracing records, and multi-algorithm comparison reports, and packaged into a visual evidence package.

2. The intelligent copyright protection and infringement identification method for student innovative works according to claim 1, characterized in that, The multi-dimensional feature fusion described in step 1 employs a feature distillation strategy: a lightweight Transformer model is used to compress the BERT semantic encoding and Word2Vec word vector fusion features of the text class; a graph neural network (GNN) is used to extract the structured embedding vectors of the abstract syntax tree (AST) and control flow graph (CFG) of the code class; and a generative adversarial network (GAN) is used to optimize the CNN features and SIFT feature point fusion vectors of the visual class, forming a low-dimensional, high-recognition digital fingerprint.

3. The intelligent copyright protection and infringement identification method for student innovative works according to claim 2, characterized in that, Step 1, which involves feature extraction for textual works, also includes: identifying the subject-verb-object structure of sentences through semantic role labeling, and distinguishing between legitimate citations and malicious plagiarism by combining citation labeling detection rules, which include identifying reference identifiers, quotation marks, and source description text.

4. The intelligent copyright protection and infringement identification method for student innovative works according to claim 2, characterized in that, Step 1, which involves feature extraction for code-based works, also includes: normalizing variable names and function names in the code, extracting execution path features of algorithm logic segments, and generating a digital fingerprint resistant to code formatting tampering by combining the semantic encoding features of code comments.

5. The intelligent copyright protection and infringement identification method for student innovative works according to claim 4, characterized in that, Step 1, which involves feature extraction for visual works, also includes: fusing color histograms, texture features, and deep learning features; performing multi-scale scaling and rotation invariance processing on the image; and using feature point matching to resist cropping, color correction, and watermark overlay infringement and tampering methods.

6. The intelligent copyright protection and infringement identification method for student innovative works according to claim 1, characterized in that, The blockchain evidence storage described in step 1 adopts a time-series hash chain plus zero-knowledge proof mode: the digital fingerprints of the first draft, revised version and final version of the work are linked together in time sequence to form a hash chain, and only the hash chain summary and metadata of key creation nodes are uploaded to the consortium blockchain. The validity of the evidence storage is verified by zero-knowledge proof without disclosing the content of the work.

7. The intelligent copyright protection and infringement identification method for student innovative works according to claim 1, characterized in that, The deep verification layer described in step 3 also includes a cross-modal comparison mechanism: by using a cross-modal attention mechanism, the feature vectors of different types of works are mapped to the same semantic space, and the structural correlation between text and code, design drawings and physical descriptions is calculated to achieve cross-type infringement identification.

8. The intelligent copyright protection and infringement identification method for student innovative works according to claim 1, characterized in that, The creation trajectory recording in step 2 also includes: generating a feature evolution curve based on the work modification operation log, calculating the creation credibility score by comparing the feature differences between the draft and the final version, and the score is used to dynamically adjust the infringement determination threshold in step 3.

9. The intelligent copyright protection and infringement identification method for student innovative works according to claim 1, characterized in that, It also includes steps for defining the originality of AI-assisted creation: by using a comparative learning model to identify the proportion of AI-generated content in the work, and combining the student's modification traces and logical integration evidence, the original contribution is calculated, and the copyright ownership boundaries of human-computer collaborative works are clarified.

10. The intelligent copyright protection and infringement identification method for student innovative works according to claim 1, characterized in that, The visualization evidence package mentioned in step 4 includes: a feature overlap heatmap, a comparison video of the creation timeline, and a weighted table of multi-algorithm judgment results.

Images