A course arrangement structure-based education content steganographic identification method and system

Abstract

This invention discloses a method, extraction method, and system for embedding and extracting steganographic identifiers of educational content based on curriculum arrangement structure, belonging to the application of information steganography and digital rights protection in the field of educational informatization. The embedding method constructs a structured candidate library containing multiple arrangement dimensions for target knowledge points. After controlled teaching effect comparison tests and confirmation of no significant differences using independent samples t-tests, candidate content items are included in a replaceable candidate set. The steganographic identifiers are converted into bit sequences. Target candidate content items are selected from the candidate sets of each arrangement dimension for encoding, generating the target teaching content. The combination patterns of candidate content items in each dimension have a definite mapping relationship with the bit sequences. The extraction method performs structural analysis on the teaching content and matches it with the candidate library, recovering the bit sequences to extract the steganographic identifiers. This invention encodes in a discrete option space within the curriculum arrangement structure, ensuring that the identifiers are not lost due to text rewriting.

Description

Technical Field

[0001] This invention relates to the intersection of information security and educational informatization, specifically to a method and system for encoding traceability identifiers into the curriculum structure of teaching content, belonging to the application of information steganography and digital rights protection technology in the field of educational informatization. Background Technology

[0002] With the large-scale promotion of national smart education platforms, copyright traceability of AI-generated educational content has become a core security requirement for educational informatization. As AI-generated educational content (AIGC-EDU) is widely applied, the need for copyright traceability of generated content is becoming increasingly urgent. In the education sector, an AI-generated math lecture, an AI-generated English speaking exercise, or a set of AI-generated programming exercises can all be copied and disseminated without authorization, resulting in damage to the intellectual property rights of content providers and educational institutions. However, existing text watermarking technologies have the following fundamental flaws:

[0003] First, existing text watermarking schemes (such as the red-green list scheme used by Kirchenbauer et al., "A Watermark for Large Language Models", ICML 2023) operate at the word level, embedding watermarks by influencing the token sampling probability of large language models. When the text is rewritten or replaced with synonyms, the watermark information is lost. In educational content scenarios, students or institutions can easily circumvent word-level watermarks by simply rewriting the handouts, rendering copyright tracing ineffective.

[0004] Second, the synonym replacement watermarking scheme encodes information by selecting specific words from synonym groups. However, it has limited robustness when large language models rewrite the entire text, which is particularly evident in scenarios where educational content is rewritten in batches.

[0005] Third, all the above solutions operate on the surface form of the text, failing to utilize the deep curriculum structure information unique to educational content. Within educational content, the same knowledge point can be taught through different equivalent examples, equivalent problem-solving paths, and equivalent analogies. While these teaching choices do not differ significantly in teaching effectiveness, their combination patterns contain a rich information encoding space—a structural feature that text rewriting attacks cannot eliminate.

[0006] Fourth, in the field of structured watermarking, Hou et al. (NAACL 2024, arXiv:2310.03991, i.e., the SemStamp scheme) proposed a sentence-level semantic watermarking based on Locality Sensitive Hashing (LSH). This method embeds the watermark by partitioning the sentence's semantic embedding space and rejecting sampling, exhibiting a certain degree of robustness against rewriting attacks. Furthermore, Yang et al. (AAAI2022) proposed a natural language watermarking based on BERT context-aware lexical substitution, encoding information through word-level semantic equivalence substitution. The common limitation of these schemes is that their operational granularity remains at the word or sentence level text surface; equivalence verification relies on automatic evaluation by semantic similarity models rather than statistical verification of teaching effectiveness; and neither is suitable for educational content scenarios with strict constraints on teaching quality.

[0007] The core difference between this invention and the aforementioned prior art lies in: (1) the granularity of operation is raised from the word / sentence level text surface to the course arrangement structure layer (discrete teaching decision level such as example selection, solution path selection, analogy method selection, etc.); (2) the equivalence verification is upgraded from automatic semantic similarity evaluation to a statistical testing mechanism based on the consistency verification of teaching effect (in the preferred embodiment, the independent sample t test is used, p>0.05); (3) the course arrangement structure selection mode does not change due to text rewriting, ensuring the indelibility of steganographic markers from a structural level. Specifically, the LSH partitioning mechanism of D2 (SemStamp) has a bidirectional semantic drift problem in the paragraph-level rewriting scenario: on the one hand, a large-scale rewriting may cause the sentence semantic vector to drift across the partition boundary, resulting in the loss of watermark information (affecting reliability); on the other hand, a small-scale rewriting may leave the semantic vector in the original partition, allowing attackers to replace the text content without destroying the watermark (affecting security). This bidirectionality stems from the inherent properties of continuous semantic embedding spaces: in a continuous space, any rewriting operation corresponds to a continuous offset of the semantic vector, and the magnitude of the offset determines whether it crosses the partition boundary. Attackers have a certain degree of uncertainty control over this. In contrast, the course arrangement option space of this invention is a nominal-scale discrete space—there is no concept of distance between candidate content items, and changing any arrangement dimension of the option will inevitably produce an observable change in the teaching content structure, which is independent of the magnitude of the rewriting operation (see

[0017] for details).

[0008] Furthermore, a significant interdisciplinary knowledge gap exists between the existing fields of text watermarking (D1, D2, D3) and educational informatization, making the technical solution of implementing steganography at the curriculum arrangement structure layer non-obvious to those skilled in the art: First, the concept of "equivalent unit" for those skilled in text watermarking is limited to semantic equivalence of words or semantic similarity of sentences, and does not involve the concept of "teaching equivalence" in pedagogy; Second, those skilled in the field of instructional content design do not have the technical background of information steganography and will not regard curriculum arrangement decisions as steganography carriers; Third, starting from the continuous semantic embedding space of D2, the natural technical path along the direction of granularity improvement is still a continuous space operation. Transforming the technical path from "improving the granularity of operation" to "changing the mathematical properties of the operation space (continuous → discrete)" requires a framework transformation, which is not a conventional technical choice. Summary of the Invention

[0009] The present invention addresses the following key technical problems: (1) how to construct a structured candidate library for the course arrangement structure of educational content and form a replaceable candidate set under multiple arrangement dimensions, so that it has structural robustness against text rewriting attacks; (2) how to verify the consistency of teaching effect of candidate content items under the premise of ensuring that teaching quality is not reduced; (3) how to encode steganographic identifiers into multiple arrangement dimensions and achieve robust encoding and decoding and fault-tolerant recovery. Technical solution

[0010] This invention proposes a method for embedding and extracting steganographic identifiers in educational content based on curriculum arrangement structure. By constructing a multi-dimensional set of replaceable candidates within the curriculum arrangement structure of the teaching content, the steganographic identifiers are encoded into the candidate selection patterns of the arrangement structure, rather than relying on word-level synonym substitution, fundamentally improving the identifiers' resistance to text rewriting attacks. The embedding method is responsible for encoding the steganographic identifiers into the arrangement structure of the teaching content, while the extraction method is responsible for recovering the steganographic identifiers from the teaching content to be tested. Both methods can be implemented independently by different entities.

[0011] (a) Definition of Arrangement Dimensions. The "arrangement dimension" referred to in this invention refers to the content organization decision-making dimension directly related to the way knowledge points are conveyed in the teaching content. That is, it refers to the discrete teaching choices that can be made for the same knowledge point during the instructional design process, affecting the learner's cognitive path, excluding purely formal layout dimensions such as fonts and page layouts. Multiple arrangement dimensions are identified in the educational content: example selection dimension (selecting specific example combinations from a candidate example library to encode information), problem-solving path selection dimension (selecting a specific path from multiple equivalent problem-solving paths for the same question), scaffolding order dimension (selecting a specific order from a predefined set of equivalent teaching scaffolding orders, each predefined order corresponding to a candidate content item, encoded with log2(M) bits of information, where M is the number of predefined orders), and analogy selection dimension (selecting a specific analogy from an equivalent analogy library). Furthermore, in language subject teaching content, a contextual scenario dimension can also be set (selecting specific teaching contextual scenarios from an equivalent contextual scenario library, such as campus scenarios, family scenarios, travel scenarios, etc.). Each arrangement dimension can encode log2(N) bits of information, where N is the number of candidate content items for that dimension. The above dimensions cover the core decision-making dimensions for the arrangement of teaching content and are applicable to structured teaching content in subjects such as mathematics, Chinese, English, physics, and chemistry.

[0012] (b) Construction of a structured candidate library. For each knowledge point in the knowledge graph, candidate content items are constructed or obtained under each arrangement dimension. The grouping of candidate content items is based on consistency verification—any candidate content item in the same replaceable candidate set shows no statistically significant difference in a preset evaluation indicator (the preset evaluation indicator includes at least one of learner answer accuracy and improvement in knowledge point mastery). Consistency verification can be performed using automated semantic similarity verification or teaching effectiveness consistency verification. In the preferred teaching effectiveness consistency verification scheme, the specific verification process is as follows: (1) Candidate content items are jointly generated by domain experts and AI-assisted system. The AI-assisted system adopts a teaching content generation module based on a large language model, which is responsible for batch generating candidate content items and conducting preliminary semantic equivalence screening (specifically, the large language model is input with the knowledge point name and teaching objective as prompt words to obtain candidate content, and the cosine similarity of the text semantic vector is not less than 0.85 as the preliminary equivalence screening threshold). Experts are responsible for the final teaching rationality review; (2) Controlled teaching effectiveness comparison test (A / B test) is conducted on each group of candidate content items, and the test results are statistically analyzed using independent samples t test; (3) Only when the difference in teaching effectiveness between any two candidate content items in the group does not reach the statistical significance level (independent samples t test p>0.05) is the group included in the replaceable candidate set.

[0013] The above A / B teaching effectiveness test is the preferred verification scheme for the construction of the structured candidate library. In actual deployment, the construction and use of the structured candidate library supports the following alternative and simplified paths: (1) Automated preliminary verification alternative scheme: When it is not possible to organize real teaching experiments immediately, the semantic similarity model can be used to conduct an automated preliminary consistency assessment of the candidate content items (with the text semantic vector cosine similarity not less than 0.85 and the teaching target keyword coverage not less than 0.90 as the preliminary consistency judgment threshold), and the candidate content items that pass the automated assessment are marked as "candidates to be verified" and used for steganographic embedding. After accumulating sufficient teaching data, the formal verification of the A / B test and the candidate library update will be carried out; (2) Pre-built standardized candidate library: The structured candidate library can be pre-built by professional educational content institutions and provided to the implementer in the form of a standardized database product. The implementer does not need to conduct A / B teaching experiments on their own to directly implement the test. The verified candidate library is used for steganographic encoding. The pre-built database is indexed by subject, grade and knowledge point, and supports on-demand expansion. Implementers can obtain it through the following means: purchasing standardized candidate library products that have been verified for teaching effectiveness from professional educational content institutions, downloading and verifying candidate libraries contributed by the community from open source educational resource platforms, obtaining candidate content items from online knowledge graph platforms through API interfaces and incorporating them into the candidate library after consistency verification according to the method of this invention, or exporting existing structured teaching resources from existing teaching management systems and formatting them into the candidate library; (3) Progressive verification strategy: Implementers can start the system operation with the automated preliminary verification results first, and gradually complete the A / B test verification by collecting the actual learning data of learners during the system operation, and dynamically update the verification status of each candidate group in the candidate library. The above alternative paths ensure that those skilled in the art (including those skilled in the fields of information security and educational informatization) can implement this invention under reasonable conditions. It should be noted that the "consistency verification" mentioned in the claims covers all the above-mentioned alternative paths, including automated semantic similarity verification, A / B testing verification of teaching effectiveness, direct use of pre-built candidate libraries, and progressive verification. All alternative paths fall within the protection scope of the claims as defined by "confirmation through consistency verification that there is no statistically significant difference in the preset evaluation indicators".

[0014] (c) Joint Encoding and Verification. The bit sequence is allocated to each orchestration dimension, a check code (e.g., CRC-16 in this embodiment) is attached, and Reed-Solomon erasure coding is used to achieve fault tolerance. The encoding process is as follows: (1) The steganographic identifier to be embedded (defined in this method as a short code identifier of no more than 28 bits, hereinafter referred to as "traceability short code") is converted into a bit sequence; (2) The bit sequence is allocated in segments according to the capacity of each orchestration dimension; (3) In each orchestration dimension, the corresponding candidate content item is selected according to the allocated bits; (4) A CRC-16 check code is attached. During decoding, the steganographic identifier information is deduced by identifying the course structure selection mode in the generated content, and CRC verification is performed to confirm the integrity. When some orchestration dimensions are damaged, the Reed-Solomon erasure coding mechanism can achieve partial recovery. The fault tolerance rate is 25% under the four-dimensional configuration (that is, at most one dimension is allowed to be completely damaged but the complete steganographic identifier can still be recovered). The hierarchical relationship between CRC-16 checksum and Reed-Solomon erasure coding in the encoding and decoding process is as follows: In the encoding stage, Reed-Solomon encoding is first applied to the source short code bit sequence to generate redundant codewords (using GF(2)). 8 A finite field with primitive polynomial x 8 +x 4 +x³+x²+1 (i.e., 0x11D), the codeword length is 4 bytes under the four-dimensional configuration, of which 3 bytes are data symbols and 1 byte is redundant symbols, corresponding to a 1 / 4 fault tolerance rate. Encoding and decoding can be implemented using the standard Reed-Solomon library (such as Python's reedsolo library or an equivalent implementation). Then, CRC-16 check bits are added to each dimension bit and embedded into the corresponding arranged dimension. In the decoding stage, bits are first extracted from each dimension and CRC-16 check is performed to identify the damaged dimension. Then, the complete dimension data and the dimension mark of the CRC check failure are sent to the Reed-Solomon decoder for erasure decoding. The final traceability short code is output only when all dimension checks pass; otherwise, the erasure recovery process is triggered.

[0015] (d) Steganography Extraction Method. The extraction end performs the following structural analysis process on the teaching content to be tested: For each arrangement dimension, extract the corresponding structural features (example text, solution step features, teaching sequence, analogy description text, etc.) from the teaching content, and match the extracted features with candidate content items in the structured candidate library. The matching methods include the following three types: template matching (comparing the extracted structural features with standard templates in the candidate library), index lookup (directly locating in the candidate library index based on the key attributes of the structural features), and rule recognition (determining the category of candidate content items based on predefined feature recognition rules). For arrangement dimensions that match successfully, the bit segment carried by the dimension is recovered based on the candidate identifier of the matched candidate content item; for arrangement dimensions that fail to match, the dimension is marked as an erasure symbol. After concatenating the recovered bit segments from each dimension, a CRC check is performed. If the check passes, the steganographic identifier is directly output. If the check fails and a dimension is marked as an erasure symbol, the data from each dimension (including the erasure symbol) is sent to an erasure decoder for erasure decoding. After recovery, a CRC check is performed again to confirm the integrity of the steganographic identifier. The computational complexity of the above extraction process depends only on the number of arranged dimensions and the number of candidate content items in each dimension (all finite small integers). It does not require word-by-word analysis of the teaching text, thus its extraction efficiency is significantly higher than existing text watermark extraction schemes based on word-level statistical analysis.

[0016] Based on the three-step method for determining inventiveness as stipulated in Section 4.2 of Chapter 4, Part II of the Patent Examination Guidelines, the inventiveness analysis of this invention is as follows. Step 1: Identify the closest prior art. The closest prior art is D2 (SemStamp, Hou et al., NAACL 2024), which employs Local Sensitive Hash (LSH) partitioning in the sentence semantic embedding space and embeds watermarks using rejection sampling. Compared to D1, it offers improved robustness and is closest to the technical objective of this invention in "improving the survival rate of identifiers under text rewriting attacks." Step 2: Identify the distinguishing technical features and the actual technical problem solved by the invention. The present invention differs from D2 in three technical features: (i) the steganography basis is replaced from a continuous vector space of sentence semantic embedding to a finite discrete option space of curriculum arrangement structure (the nominal scale selection set of teaching decisions such as examples, solution paths, analogies, scaffolding order, etc.); (ii) the equivalence verification standard is replaced from unsupervised automatic judgment based on LSH semantic distance to a statistical testing mechanism based on the consistency verification of teaching effectiveness, and the verification dimension is shifted from semantic similarity to the statistical significance evaluation of teaching effectiveness (in the preferred embodiment, a controlled A / B teaching effectiveness comparison test and an independent samples t-test are used, p>0.05); (iii) a "measurability constraint of teaching effectiveness" is introduced—any change in the selection of curriculum arrangement leads to a measurable change in the learner's cognitive path, thereby preventing attackers from modifying the steganography carrier without causing a noticeable loss of teaching effectiveness. The technical problem actually solved by the above-mentioned distinguishing technical features is: "How to construct an arrangement dimension with discrete structure that is not eliminable under the constraint of teaching quality, while ensuring that the embedding of steganography markers does not reduce the quality of teaching content." The third step is to determine non-obviousness. Those skilled in the art (in the field of text watermarking / NLP) will not obviously derive the technical solution of the present invention from D2 when faced with the above-mentioned technical problems, for reasons detailed in

[0017] .

[0017] The core argument for non-obviousness lies in the bidirectional semantic drift problem in the D2 continuous space and the structural constraint guarantee of the discrete space in this invention. The robustness analysis of D2 (SemStamp) requires distinguishing between two scenarios: First, a significant rewrite causes the sentence semantic vector to drift across the LSH partition boundary, resulting in watermark bit flipping and a decreased survival rate—experimental data shows that the D2 survival rate is only 61.7% under a 50% rewrite perturbation, indicating that this scenario significantly exists. Second, a smaller rewrite leaves the semantic vector within the original partition, preserving the watermark, but the text content has been substantially altered. The existence of scenario two means that if an attacker can control the rewrite magnitude to keep the semantic drift below the partition radius, they can replace the surface content of the text without destroying the watermark; however, if the rewrite magnitude exceeds the partition radius, the watermark is destroyed. This property of "uncontrollable rewrite magnitude leading to unstable watermark survival rate" is an inherent limitation of the continuous semantic embedding space and is unrelated to the choice of partition granularity. Conversely, the course arrangement option space of this invention has the following structural constraints: the course arrangement selection is a discrete variable of the nominal scale—there is no distance concept between “using example A” and “using example B”, only “same” and “different”. Any operation that attempts to change the steganographic identifier must replace example A with example B. This operation is an observable discrete jump at the teaching level, specifically manifested as: (1) the knowledge point coverage structure changes (different examples activate different knowledge point paths); (2) the learner's cognitive load distribution changes (different examples have different difficulty gradients); (3) under the teaching effect consistency verification standard of the replaceable candidate set, changing the candidate content item means exiting the “replaceable” set, which leads to measurable differences in teaching effect. These three points make it logically impossible for an attacker to change the steganographic carrier without affecting the integrity of the teaching structure, which is in stark contrast to the uncertainty of “the rewriting range determines the fate of the watermark” in continuous space.

[0018] The technical features that distinguish this invention from the two closest prior art technologies, and the lack of technical inspiration, are analyzed as follows. The difference from D1 (Kirchenbauer word-level red-green list scheme): D1 operates on the token sampling probability distribution of LLM, its technical framework is built within the lexical-semantic space, it does not involve the structural layer of teaching content, and it does not provide any technical inspiration for using curriculum arrangement decisions as a steganographic encoding carrier. The differences between D2 (SemStamp sentence-level semantic watermarking scheme) and D2 are as follows: (i) The discreteness of the operation space is shifted—D2, when extended along the direction of granularity improvement, still belongs to the continuous semantic embedding space operation. The leap from the continuous semantic embedding space to the discrete teaching decision option space requires the introduction of the concept of curriculum arrangement equivalence in the field of education. D2 does not provide technical inspiration for this leap; (ii) The subject-specific leap of equivalence verification—D2 uses LSH partitioning as the automatic equivalence judgment standard. This invention leaps the equivalence verification to the level of teaching effect statistical testing. The concept of teaching effect does not exist in the technical framework of D2; (iii) The structural guarantee of attack impossibility—D2's robustness depends on the LSH partitioning granularity adjustment. The robustness of this invention comes from the discrete and non-eliminable nature of curriculum arrangement selection (see

[0017] ), which is a qualitative difference rather than a quantitative improvement. In summary, neither D1 nor D2 provides technical inspiration for steganographic encoding in the discrete teaching decision space of curriculum arrangement structure. This invention is non-obvious.

[0019] Explanation of the correspondence between the technical solutions and inventive arguments of the independent claims. In independent claim 1 (embedding method), S1 (constructing or obtaining a structured candidate library with multiple arrangement dimensions) corresponds to the distinguishing technical feature (i) – the nominal scale discreteness of course arrangement selection; S2 (after confirming no statistically significant difference through consistency verification, it is included in the set of replaceable candidates) corresponds to the distinguishing technical feature (ii) – the consistency verification mechanism; step S2 of independent claim 1 has already limited the verification method to consistency verification of teaching effectiveness (controlled A / B teaching effectiveness comparison test and independent sample t test), and dependent claim 3 further limits the statistical significance level α=0.05 and the test sample size to no less than 30 learners, providing a fallback space for specific statistical parameters. S4 (selecting target candidate content items in each arrangement dimension based on the bit sequence) combined with S5 (generating target teaching content containing structural features, wherein the structural features include the combination pattern of the selected target candidate content items in each arrangement dimension, and the combination pattern has a definite mapping relationship with the bit sequence) and the definition of discrete arrangement dimensions in S1, together support the distinguishing technical feature (iii) – the guarantee of the impossibility of attack on discrete arrangement dimensions. Independent claim 6 (extraction method) independently protects the extraction process of steganographic identifiers, covering the implementation behavior of the embedding party and the extraction party, respectively, along with claim 1. The aforementioned distinguishing technical features are explicitly defined in the independent claims, working together to solve the practical technical problem of "countering text rewriting attacks under the constraint of teaching quality." Dependent claims 4 (checksum) and 5 (error correction code) further characterize the beneficial effects of the present invention compared to the prior art (D1, D2) from the two dimensions of integrity and robustness, respectively, and are corroborated by the experimental data in

[0021] and

[0022] . Beneficial effects

[0020] The technical solution provided by this invention has the following beneficial effects:

[0021] (1) In the prototype system experimental verification (experimental environment: NVIDIA A100 40GB GPU×1, operating system Ubuntu 22.04, Python 3.10), the test dataset contains 200 AI-generated teaching contents for two subjects, mathematics and English (100 contents per subject, text length 300 to 800 words, generation model is GPT-4-turbo, generation temperature parameter temperature=0.7, top_p=0.95). The text rewriting attack uses a large language model (GPT-4-turbo, rewriting instruction is "under the premise of retaining all knowledge points and teaching structure, rewrite the following teaching contents at the paragraph level, requiring the rewritten content to have no more than 50% overlap with the original text", temperature=0.8) to rewrite the original text at the paragraph level, retaining the knowledge point coverage structure but completely rewriting the word expression. Under the four-dimensional configuration, the average survival rate of the steganalysis marker under the 50% text rewriting perturbation condition reached 85.2% (confidence interval 82.1% to 88.3%). The above experimental data are the results of the prototype system verification. Actual deployment performance depends on the size of the candidate library and the orchestration dimension configuration. Compared to the baseline: Kirchenbauer's word-level red-green list watermarking scheme has a survival rate of only 45.3% under the same perturbation, while the synonym replacement watermarking scheme has a survival rate of 52.1%. This method improves upon the word-level scheme by 40 percentage points (two-sample z-test p<0.001). Regarding the efficiency of steganography extraction, this method only requires analyzing the course structure selection pattern (discrete candidate content item identification), with a single-article extraction time of 0.3 seconds. Compared to Kirchenbauer's full-text word-by-word statistical analysis, which takes 2.1 seconds, the extraction efficiency is improved by 7 times.

[0022] In a supplementary comparative experiment with the D2 (SemStamp, Hou et al. NAACL 2024) baseline, under the same experimental environment and 50% text rewriting perturbation, the mean watermark survival rate of the SemStamp scheme was 61.7% (confidence interval 58.2% to 65.2%). The survival rate of the four-dimensional configuration of the steganographic identifier in our method (85.2%) was 23.5 percentage points higher than that of SemStamp (two-sample z-test p<0.001). The LSH partition of SemStamp had a higher survival rate under paragraph-level rewriting than the word-level scheme (D1, 45.3%), but it was still significantly lower than that of our method, verifying the structural advantage of the discrete replaceable candidate set space of the course arrangement structure compared with the continuous space of sentence-level semantic embedding in resisting rewriting attacks.

[0023] In supplementary experiments using a two-dimensional configuration (using only example selection and analogy selection dimensions), the mean steganalysis survival rate reached 71.4% (confidence interval 67.8% to 75.0%) under 50% text rewriting perturbation, still significantly better than the word-level baseline scheme (45.3%). The two-dimensional configuration is suitable for lightweight deployment scenarios with relatively simple knowledge point structures and small candidate library sizes, such as comprehension exercises in lower elementary Chinese textbooks. These experimental data demonstrate that our method provides superior steganalysis survival rate guarantees compared to existing technologies under configurations with at least two arrangement dimensions.

[0024] (2) Under the standard deployment configuration where the number of candidate content items in each dimension is no less than 128, the average embedding of a single text in the four dimensions is 28.6 bits, which is sufficient to carry a 28-bit short code traceability identifier; under the two-dimensional configuration, the average embedding is about 14.3 bits, which can carry a 16-bit short code traceability identifier. Steganography has no negative impact on the teaching quality of the generated content (the differences between the teaching effect comparison test groups in each embodiment did not reach the statistical significance level). Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall process of the course structure steganography identification method of the present invention, which shows the complete process from knowledge graph knowledge points, through the construction of structured candidate library, bit sequence allocation, selection of arrangement dimension candidates to the generation of teaching content.

[0026] Figure 2 A detailed schematic diagram of the multi-dimensional steganography encoding and steganography identifier embedding / extraction process is provided, demonstrating the parallel encoding mechanism of four structural arrangement dimensions, the CRC-16 check appending process, and the Reed-Solomon erasure coding fault tolerance and recovery mechanism. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The present invention is applicable to copyright tracing of various types of structured educational content; typical embodiments for mathematics and English are given below.

[0028] Step 1 (This embodiment is a minimal feasibility verification configuration, only used to verify the feasibility of the method. In actual deployment, the number of candidate content items in each dimension can be expanded to dozens or even hundreds to significantly improve the embedding capacity): Construct a structured candidate library for the knowledge point

Solution of quadratic equations in one variable

[0029] The parameters for verifying the consistency of teaching effectiveness in the structured candidate library construction (the following is a preferred configuration in this embodiment, which implementers can adjust reasonably according to actual conditions (number of available learners, testing period constraints, etc.)) are as follows: the test sample size for each group of candidate content items is no less than 30 junior high school learners (preferred configuration), the test lasts for 2 weeks (including the complete process of pre-test-learning-post-test), and the subjects are students of a certain middle school mathematics course. In the preferred verification scheme of this embodiment, the difference in teaching effectiveness between any two candidate content items within a group is tested using an independent samples t-test, with a significance level of α=0.05. The checksum adopts the CRC-16 standard.

[0030] Step 2: Assign the first 5 bits of the steganographic identifier source code as follows: Example dimension code "01" (select the second problem x²+5x+6=0), Solution dimension code "10" (select matching method), Analogy dimension code "1" (select decryption game), Scaffolding dimension code "1" (select example problem-definition introduction-practice order).

[0031] Step 3: When generating teaching content, generate corresponding teaching segments based on the selected candidate content items to complete the steganography embedding. The generated content includes complete teaching text containing the selected examples, solutions, and analogies. The teaching effectiveness is not statistically significantly different from the version without steganography embedding.

[0032] Step 4: Steganography identification. The following structural analysis process is performed on the teaching content to be tested: (1) Example identification: Extract the mathematical problem text in the teaching content and perform template matching with each candidate example in the candidate example database (prioritize exact matching, and use mathematical expression structure equivalence judgment when matching fails, such as comparing the coefficient set of the equation {1,3,2} with the coefficient set of each candidate equation), and output the matching candidate example number or mark it as an erasure symbol; (2) Solution identification: Analyze the key steps of the solution process in the teaching content (such as whether the characteristic phrases such as "factorization", "substitution into formula", and "completing the square" appear). (or the corresponding mathematical operation mode), and perform feature matching with each candidate solution in the candidate solution library through rule recognition, and output the matching candidate solution number; (3) Analogy recognition: extract the analogy description text in the teaching content, and perform index search matching with each candidate analogy in the candidate analogy library, and output the matching candidate analogy number; (4) Scaffolding sequence recognition: analyze the paragraph structure of the teaching content, locate the appearance order of teaching links such as "definition introduction", "example", "exercise" through rule recognition, compare with the predefined order in the scaffolding sequence library, and output the matching sequence number. The recognition results of each dimension are mapped back to the corresponding bits according to the candidate identifier, and the CRC-16 verification is performed after splicing the bit sequence of each dimension. If the verification is successful, the steganographic identifier is successfully extracted; if the verification fails, the unrecognizable dimension is marked as an erasure symbol and sent to the Reed-Solomon decoder to try to erase and restore.

[0033] Step 1: Construct a structured candidate database for the knowledge point

English Vocabulary - Adjectives Expressing Emotions (Level B1)

[0034] The parameters for verifying the consistency of teaching effectiveness in the structured candidate library (preferred configuration, which can be adjusted by the implementer according to actual conditions) are as follows: the test sample size is no less than 30 junior high school English learners, and the test lasts for 2 weeks (including the complete process of pre-test - learning - post-test). The differences in teaching effectiveness among the candidate content items in each group did not reach a statistically significant level, meeting the conditions for inclusion in the replaceable candidate set.

[0035] Step 2: Assign the first 5 bits of the steganographic identifier source code to the following dimensions: Example sentence dimension code "10" (select She felt excited about the trip.), Analogy dimension code "0" (select the mood palette analogy), Context dimension code "10" (select the travel scenario), and Scaffolding dimension code "1" (select the scenario practice - example sentence - word meaning explanation order).

[0036] Step 3: When generating English vocabulary teaching content, generate corresponding teaching segments according to the selected candidate content items and complete the embedding of steganography markers.

[0037] Step 4: Steganography identification. The English teaching content to be tested is subjected to structural analysis: (1) Example sentence identification: Extract English example sentences from the teaching content and perform template matching with each candidate example sentence in the candidate example sentence library (prioritize exact matching, and use syntactic structure + keyword combination matching when matching fails), and output the matching candidate example sentence number; (2) Analogy identification: Extract the lexical analogy descriptions from the teaching content and perform index search matching with the candidate analogy library; (3) Context scene identification: Analyze the context description keywords in the teaching content (such as scene identifier words such as "campus", "family", "travel", etc.), and compare them with the candidate context library through rule identification; (4) Scaffolding order identification: Analyze the order of appearance of teaching links (the order of "word meaning explanation", "example sentence", "scenario practice", etc.) through rule identification and compare them with the predefined order library. The identification results of each dimension are mapped back to the bit sequence, spliced ​​and then checked by CRC-16. When the check fails, the unrecognizable dimension is marked as an erasure symbol and sent to the erasure decoder for recovery.

[0038] Step 1: Construct a structured candidate library for the knowledge point

Newton's Second Law F=ma

[0039] Consistency verification parameters for teaching effectiveness in the structured candidate library (preferred configuration): The test sample size is no less than 30 high school physics learners, and the test lasts for 2 weeks (including the complete process of pre-test - learning - post-test). The differences in teaching effectiveness among the candidate content items in each group did not reach a statistically significant level, meeting the conditions for inclusion in the replaceable candidate set.

[0040] Step 2: Assign the first 5 bits of the steganographic identifier source code to the following dimensions: Example dimension code "11" (select the elevator weighing change problem), Solution dimension code "01" (select the isolation method), Analogy dimension code "0" (select the analogy between accelerator and car acceleration), and Scaffolding dimension code "1" (select the experimental observation - pattern summary - example consolidation order).

[0041] Step 3: When generating physics teaching content, generate corresponding teaching segments according to the selected candidate content items and complete the embedding of steganographic identifiers.

[0042] Step 4: Steganography extraction. Perform structural analysis on the physics teaching content to be tested: (1) Example identification: Extract the physics problem text from the teaching content, and use physical quantity feature extraction (identify the set of physical quantities involved in the problem, such as {force, mass, acceleration, ...}). (1) Sloping angle, scene keywords such as "elevator", "sloping", "horizontal plane" and "suspended" are matched with the candidate example question bank to output the matching candidate example question number; (2) Solution identification: Analyze the solution process in the teaching content, and detect the feature pattern through rule identification (such as the feature of "overall method" is to list F=ma for the whole system without separating the forces; the feature of "isolation method" is to list the equation after analyzing the force of a single object; the feature of "orthogonal decomposition method" is to establish a coordinate system and list the equations in the x and y directions respectively), and output the matching candidate solution question number; (3) Analogy identification: Extract the analogy description text in the teaching content and perform index search matching with the candidate analogy library; (4) Scaffolding sequence identification: Analyze the appearance order of the teaching links through rule identification (the order of "law derivation", "example demonstration", "experiment verification" or "experiment observation", "law summary" and "example consolidation" links), and compare with the predefined sequence library. The identification results of each dimension are mapped back to the bit sequence, concatenated, and then checked using CRC-16. If the check fails, the unrecognizable dimension is marked as an erasure symbol and sent to the erasure decoder for recovery.

Claims

1. A method for embedding steganographic identifiers in educational content, wherein the method is implemented by a processor executing a computer program stored in memory, characterized in that, Includes the following steps: S1. Construct or obtain a structured candidate library for the target knowledge points. The structured candidate library contains candidate content items under multiple arrangement dimensions, and assigns a corresponding candidate identifier to each candidate content item. S2. For multiple candidate content items expressing the same knowledge point under the same arrangement dimension, after consistency verification confirms that there is no statistically significant difference in the preset evaluation indicators, they are classified into the same replaceable candidate set, forming at least two replaceable candidate sets; the preset evaluation indicators include at least one of learner answer accuracy rate and knowledge point mastery improvement; the consistency verification includes teaching effectiveness consistency verification, which includes: conducting controlled teaching effectiveness comparison tests on candidate content items in the same replaceable candidate set, using independent samples t-test to statistically analyze the test results, and confirming that the group of candidate content items meets the consistency verification requirements when the difference between any two candidate content items in the preset evaluation indicators does not reach the statistical significance level; S3. Convert the steganographic identifier to be embedded into a bit sequence, and determine the encoding capacity of the corresponding orchestration dimension according to the number of candidates in each alternative candidate set, where the encoding capacity of each orchestration dimension is log2(N) bits, and N is the number of candidate content items in the alternative candidate set corresponding to the orchestration dimension. S4. Based on the bit sequence, select the target candidate content item from the replaceable candidate set corresponding to each arrangement dimension to obtain the candidate selection result; S5. Generate target teaching content based on the candidate selection results, such that the arrangement structure of the target teaching content includes the structural features corresponding to the candidate selection results; the structural features include the combination pattern of the selected target candidate content items in each arrangement dimension, and the combination pattern has a definite mapping relationship with the bit sequence.

2. The method according to claim 1, characterized in that, The arrangement dimensions include at least two of the following: example selection dimension, problem-solving path selection dimension, scaffolding order dimension, analogy selection dimension, and contextual scenario dimension.

3. The method according to claim 1, characterized in that, In the controlled teaching effectiveness comparison test, the statistical significance level of the independent samples t-test is α=0.05, and the test sample size for each candidate content item is no less than 30 learners.

4. The method according to claim 1, characterized in that, After converting the steganographic identifier into a bit sequence in step S3, a checksum is also added to the bit sequence.

5. The method according to claim 1, characterized in that, After converting the steganographic identifier into a bit sequence in step S3, the step also includes error correction encoding of the bit sequence.

6. A method for extracting steganographic identifiers from educational content, wherein the method is implemented by a processor executing a computer program stored in memory, characterized in that, Includes the following steps: E1. Obtain a structured candidate library, which contains candidate content items under multiple arrangement dimensions and candidate identifiers corresponding to each candidate content item. The candidate content items are categorized into replaceable candidate sets according to the arrangement dimensions. E2. Obtain the target teaching content to be tested; E3. Perform structural analysis on the target teaching content and identify candidate content items actually used in each arrangement dimension; E4. Based on the candidate identifiers of the identified candidate content items, recover the bit segments corresponding to each arrangement dimension and splice them together to obtain the bit sequence; E5. Extract the steganalysis identifier based on the bit sequence.

7. The method according to claim 6, characterized in that, The structural analysis described in step E3 includes at least one of the following methods: template matching, index lookup, and rule recognition; for successfully matched orchestration dimensions, the bit segments carried by the orchestration dimension are recovered based on the candidate identifiers of the matched candidate content items; for unmatched orchestration dimensions, the recognition result of the orchestration dimension is marked as an erasure symbol.

8. The method according to claim 6, characterized in that, Step E5 also includes: verifying the recovered bit sequence with a check code; if the verification passes, the steganographic identifier is confirmed to have been successfully extracted; if the verification fails and there are arrangement dimensions marked as erasure symbols, the data of each dimension containing erasure symbols are sent to the erasure code decoder to perform erasure recovery, and after recovery, the check code is verified again to confirm the integrity of the steganographic identifier.

9. A system for embedding steganographic identifiers for educational content, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the method of any one of claims 1 to 5.

10. A system for extracting steganographic identifiers for educational content, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the method of any one of claims 6 to 8.

Images