A privacy data collaborative computing method in a computer trusted execution environment

Abstract

The application relates to the cross field of privacy calculation and educational information technology, and specifically discloses a privacy data collaborative calculation method in a computer trusted execution environment, which comprises the following steps: in the trusted execution environment, performing operation intention declaration and multi-modal privacy feature trusted decoupling packaging on three types of campus culture immersion data to generate a privacy protection immersion feature vector, a coupling residual vector and a verifiable certificate; based on a strategy heterogeneity subspace and a counterfactual causal inference mechanism, calculating a cross-node net attribution contribution degree and a strategy heterogeneity calibration factor, and by a distributed consensus network, blocking a violation access to generate an audit record; detecting a student digital literacy critical transition point, calculating a strategy deviation, generating a directional optimization gradient, updating a local campus culture immersion strategy model, and returning a parameter digest and a transition threshold to the distributed consensus network. The application realizes multi-path effect trusted decoupling, heterogeneity strategy calibration attribution and immersion strategy continuous evolution under the premise that original education data does not go out of the domain.

Description

Technical Field

[0001] This invention relates to the intersection of privacy computing and educational information technology, and in particular to a method for collaborative computing of privacy data in a trusted execution environment of a computer. Background Technology

[0002] With the deep integration of artificial intelligence and intelligent manufacturing technologies into vocational education in the field of mechanical engineering, the construction of campus culture in vocational colleges is undergoing a paradigm shift from equipment display to digital literacy cultivation. Mechanical engineering colleges are using three pathways—environmental culture transformation, activity culture design, and AI-assisted learning—to continuously cultivate students' digital awareness, digital skills, digital ethics, digital collaboration, and intelligent applications across five dimensions. This has generated a wealth of cross-modal and temporally coupled educational data, encompassing students' activity trajectories in cultural corridors, interaction logs from club activities, equipment operation sequences during practical training weeks, and reflection records from AI-assisted dialogues.

[0003] Existing privacy-preserving computation technologies suffer from three deep-seated flaws in this field. First, the coupling effects of multi-path cultural immersion cannot be separated. The three paths—environmental culture, activity culture, and AI agent interaction—exhibit complex nonlinear interaction effects. Existing solutions can only statistically calculate the effects of a single path or the resulting mixture, failing to separate the coupling residuals between paths from the overall effect, leading to distorted attribution results. Second, the heterogeneity of strategies among teaching nodes distorts collaborative aggregation. Cultural immersion strategies across different classes, departments, and even industry-university nodes differ significantly in content depth, implementation pace, and evaluation standards. Directly aggregating the privacy protection features of each node introduces policy conflict noise. Existing solutions lack mechanisms for quantifying and calibrating the heterogeneity of node strategies under privacy protection. Third, the nonlinear leaps in students' digital literacy development are difficult to capture. Students do not grow linearly during continuous cultural immersion but rather undergo a qualitative leap from quantitative change at specific critical points. Existing time-series analysis methods can only perform smooth trend fitting and cannot detect these critical leap points, causing the direction of strategy evolution to deviate from the actual growth pattern.

[0004] Therefore, there is an urgent need for a privacy-preserving collaborative computing method for campus culture construction in mechanical vocational education under a trusted computer execution environment. This method should ensure that the original educational data does not leave the domain, and achieve privacy-preserving decoupling and separation of multi-path cultural immersion coupling effects, trusted calibration and attribution of heterogeneity of teaching node strategies, and directional evolution of immersion strategies under the perception of critical leaps in student literacy. Summary of the Invention

[0005] The purpose of this invention is to solve the deep-seated technical problems in the construction of campus culture in mechanical vocational education, namely, the difficulty in separating the multi-path cultural immersion coupling effect, the difficulty in calibrating the heterogeneity of teaching node strategies, and the difficulty in perceiving the critical leap of students' literacy. The invention proposes a privacy data collaborative computing method in a trusted computer execution environment.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for collaborative computation of privacy-preserving data in a trusted execution environment includes the following specific steps: S1. In the scenario of campus culture construction in mechanical vocational education, each teaching node collects original multi-source campus culture immersion data in a trusted execution environment. This original multi-source campus culture immersion data includes environmental culture immersion data, activity culture immersion data, and AI-interactive immersion data. Each teaching node maps the three types of data collected to a unified educational timeline through a culture immersion time window. The collaborative computing coordinator receives the original collaborative computing request submitted by the computing task initiator. This original collaborative computing request includes a computing task identifier, a list of participants, and an executable computing program. The collaborative computing coordinator performs static code analysis on the executable computing program, extracting each instruction sequence from the executable computing program. The system extracts the target data field identifiers and corresponding opcode types for each instruction related to data access, and organizes all extracted instructions and their corresponding target data field identifiers and opcode types into a program data operation relationship diagram according to the control flow relationship of the executable computing program. The collaborative computing coordinator sends the program data operation relationship diagram to each teaching node. Each teaching node, based on its preset data usage strategy corresponding to its locally stored privacy data, reviews the opcode types and their execution order relationships related to its privacy data in the program data operation relationship diagram, and generates its own first usage declaration credential. The first usage declaration credential includes the first instruction that the teaching node is allowed to execute. The system comprises an opcode type whitelist and a second opcode type blacklist, data field identifiers corresponding to each opcode type, and the program location range where each opcode type is allowed to appear in the control flow. Each teaching node digitally signs the first purpose declaration credential using its private key and returns it to the collaborative computing coordinator. The collaborative computing coordinator performs intersection and union operations on the first opcode type whitelist and the second opcode type blacklist in the first purpose declaration credential returned by each teaching node, generating a collaborative computing task specification. The collaborative computing task specification includes an executable computing program, a unified opcode type whitelist, a unified opcode type blacklist, and the program location corresponding to each opcode type. Interval constraints; after completing the purpose declaration interaction with the collaborative computing coordinator and receiving the collaborative computing task specification, each teaching node uses a multimodal feature extraction model to generate a privacy-preserving immersion feature vector from the campus culture immersion data mapped to the unified education timeline, uses a path coupling decoupling model to generate a coupling residual vector representing the nonlinear interaction effect between the three types of immersion paths, and generates a verifiable credential based on zero-knowledge proof. The verifiable credential is used to prove that the decoupling encapsulation and purpose declaration process is completed in a trusted execution environment and has not been tampered with. The privacy-preserving immersion feature vector and the coupling residual vector are both desensitized through a differential privacy mechanism and have data purpose tags corresponding to the collaborative computing task specification attached. S2. Each teaching node publishes a privacy-protected infiltration feature vector, a coupling residual vector, and a verifiable credential, all tagged with the data usage description, to the distributed consensus network via a trusted data delivery protocol. This trusted data delivery protocol includes at least three steps: encrypted data transmission, digital signature by the sender, and confirmation receipt by the receiver. After verifying the validity of the verifiable credential, the collaborating node performs cross-domain alignment based on the temporal coupling relationship of the educational timeline and the coupling residual vector to construct a strategy heterogeneity subspace. This subspace quantifies the differences in campus culture infiltration strategies among teaching nodes. Within this strategy heterogeneity subspace, a counterfactual causal inference mechanism is used to calculate the differences in different campus culture infiltration strategies. The cross-node net attribution contribution of the cultural immersion path to students' digital literacy growth is calculated by the counterfactual causal inference mechanism, which compares the differences in student literacy growth under the actual immersion strategies adopted by each teaching node with the constructed control strategies, and generates a strategy heterogeneity calibration factor. In the attribution calculation process, the distributed consensus network blocks data field access requests initiated by collaborative nodes that exceed the scope declared in the collaborative computing task specification based on the data usage label, and generates a usage execution audit record. The usage execution audit record records the data field identifier, opcode type, and blocking reason involved in the blocked data field access request. S3. Each teaching node obtains the cross-node net attribution contribution, policy heterogeneity calibration factor, and usage execution audit record from the distributed consensus network. Within its local trusted execution environment, it invokes a pre-deployed campus culture immersion strategy model. Based on the privacy-preserving immersion feature vector and coupling residual vector, it detects the critical transition point for students' digital literacy development. The coupling residual vector characterizes the comprehensive impact of the nonlinear interaction effect between the three types of immersion paths on students' digital literacy development. It calculates the policy deviation between the policy prediction output of the campus culture immersion strategy model before and after the critical transition point and the cross-node net attribution contribution. Combining the policy heterogeneity calibration factor, it generates a directional optimization gradient and updates the local campus culture immersion strategy model. Simultaneously, it uses the updated strategy model parameter summary and transition threshold as additional payloads for a remote proof report, which is then transmitted back to the distributed consensus network via a remote proof mechanism to complete the trusted collaborative evolution of the cross-node campus culture immersion strategy.

[0007] The beneficial effects of this invention are as follows: 1. By generating a coupling residual vector through a path coupling decoupling model, the nonlinear interaction effects between the three paths of environmental culture, activity culture, and AI agent interaction are separated from the main effects. This enables subsequent attribution calculations to accurately locate the net contribution of a single path, avoid repeated calculations or omissions of synergistic effects, and achieve reliable decoupling of multi-path effects under privacy protection.

[0008] 2. By constructing a strategy heterogeneity subspace, the differences in cultural immersion strategies among different teaching nodes are quantified under privacy protection. The strategy differences are introduced as covariates into counterfactual causal inference, so that the cross-node collaborative calculation results reflect the true attribution relationship under strategy heterogeneity conditions, rather than a simple statistical average, thereby improving the credibility and interpretability of the attribution results.

[0009] 3. By detecting abrupt changes in the distribution of privacy protection immersion feature vectors on the education timeline, we can identify critical leap nodes in students' digital literacy from quantitative to qualitative change. Based on the deviation between cross-node net attribution contribution and local strategy prediction, we can generate a directional optimization gradient, ensuring that the evolution direction of the campus culture immersion strategy is highly consistent with the actual growth pattern, forming a continuous evolution of the strategy under a credible closed loop. Attached Figure Description

[0010] Figure 1 This is a flowchart of the method proposed in Embodiment 1 of the present invention; Figure 2 This is a comparison diagram of the effects of the present invention and the prior art in Example 2. Detailed Implementation

[0011] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments. Example

[0012] Please see the appendix Figure 1 A method for collaborative computation of privacy-preserving data in a trusted execution environment includes the following specific steps: S1. Collect three types of infiltration data and map them to a unified education timeline to generate a program data operation relationship diagram; review the program data operation relationship diagram, generate and sign a first-purpose declaration certificate, take the intersection of the whitelist and the union of the blacklist, and generate a collaborative computing task specification; extract privacy-protected infiltration feature vectors and coupling residual vectors, generate verifiable certificates based on zero-knowledge proofs, and perform differential privacy desensitization and add data usage labels; specifically including: S11. Generation of the data acquisition mapping and program data operation relationship diagram, specifically including: S111. Each teaching node is deployed in edge computing devices in various teaching locations on campus. Each device has a built-in trusted execution environment to ensure the secure isolation of the data acquisition process. After the node starts, it collects three types of raw campus culture immersion data in parallel through multi-source sensor interfaces: The first category is environmental cultural immersion data, collected by environmental sensors deployed in classrooms and public activity areas. This includes temperature, humidity, light intensity, noise levels, and air quality indicators, with a sampling frequency set between 0.1Hz and 1Hz, ideally 0.5Hz. The second category is activity cultural immersion data, derived from activity logs in the teaching management system. This includes attendance records, student participation time, interaction frequency, and digital work submission timestamps. The third category is AI-powered interactive immersion data, recording dialogue text, voice interactions, and behavioral patterns between students and intelligent teaching assistants and virtual tutors.

[0013] After data collection, the node segments the three types of data according to a preset cultural immersion time window. The length of this time window ranges from 5 to 30 minutes, with a preferred value of 15 minutes. For timestamps from different data sources within the same window, the node converts them to Coordinated Universal Time (UTC) and maps them to a unified educational timeline according to their chronological order. If data is missing within a time window, the data missing rate is calculated, i.e.: In the formula: For data missing rate, The theoretical number of data points within this time window is determined by the product of the sampling frequency and the time window length. This represents the actual number of data points collected within that time window; when the missing rate... When the missing data exceeds the upper limit of 5% to 20%, the window is marked as abnormal. The preferred missing rate threshold is 10%. Missing data within the threshold are filled in using linear interpolation or forward imputation, ultimately forming a multi-source data sequence arranged in chronological order.

[0014] S112. The collaborative computing coordinator is deployed in the campus data center or cloud server, serving as the central scheduling entity for cross-node collaborative computing tasks. When the initiator of a computing task submits an original collaborative computing request, the coordinator receives the request through a secure transmission channel. The request is encapsulated in a structured data format, containing three core fields: The first field is the computing task identifier, using a globally unique identifier format, with a length ranging from 16 bytes to 64 bytes, preferably 32 bytes, used to uniquely identify this collaborative computing task. The second field is the list of participants, recording the unique identifiers of each teaching node in a list format, with the list length (i.e., the number of nodes) ranging from 2 to 50, preferably 5 to 10. The third field is the executable computing program, encapsulated in bytecode or container image form, with a program size ranging from 1 megabyte to 500 megabytes, preferably 10 megabytes to 50 megabytes.

[0015] After receiving the request, the coordinator performs the following parsing operations in sequence: reads the computation task identifier and verifies its uniqueness to avoid conflicts with historical tasks; extracts the list of participants and verifies the network online status and trusted execution environment readiness status of each teaching node; loads the executable computation program into the isolated sandbox environment, completes static resource allocation, and provides execution objects for subsequent static code analysis.

[0016] S113. The collaborative computing coordinator performs static code analysis on the executable computing program within an isolated sandbox. This process does not trigger actual program execution; it only performs behavioral review by parsing binary instructions or intermediate representation code. The coordinator traverses the instruction sequence line by line starting from the program's entry address. The upper limit of the instruction traversal depth ranges from 1,000 to 100,000 instructions, with a preferred value of 10,000 instructions. Loops exceeding the upper limit are either expanded or truncated.

[0017] During the traversal, the coordinator identifies instruction categories related to data access, including memory read instructions, database read instructions, file system read instructions, and network data request instructions. For each identified data access instruction, the coordinator extracts the target data field identifier from the instruction operands or register references. This identifier is in string or integer index form, with a length ranging from 1 byte to 256 bytes, preferably 32 bytes. Simultaneously, the coordinator determines the corresponding opcode type, classifying it as a read opcode, write opcode, delete opcode, aggregation opcode, or transfer opcode. The data access instruction parsing confidence threshold is set between 90% and 99%, with a preferred value of 95%. Instructions below this threshold are marked for manual review, while instructions above the threshold are directly included in subsequent relation graph construction. After extraction, the coordinator binds the target data field identifier with the corresponding opcode type and stores it as an intermediate record.

[0018] S114. The collaborative computing coordinator arranges intermediate records according to the control flow relationships of the executable computing program: The coordinator identifies sequential execution segments, conditional branch segments, and loop segments in the program through a basic block partitioning algorithm, and extracts the jump relationships and execution order between each basic block. The coordinator arranges the target data field identifiers and opcode types within the same basic block in ascending order of the program counter, establishing a temporal association. For data dependencies across basic blocks, the coordinator traces the transmission path of field identifiers between registers and memory through data flow analysis, connecting the operations of the source basic block and the operations of the target basic block with directed edges.

[0019] After orchestration, the coordinator generates a program data operation relationship graph. This graph uses data field identifiers as nodes, opcode types as edge attributes, and execution frequency as edge weights. The number of nodes ranges from 10 to 10,000, with a preferred range of 100 to 1,000; the number of edges ranges from 10 to 50,000, with a preferred range of 500 to 5,000. Once generated, the coordinator sends the program data operation relationship graph to each teaching node via a transport layer secure encrypted channel. The size of the data packet sent by a single node ranges from 1 kilobyte to 10 megabytes, with a preferred value of 100 kilobytes. Each node returns an acknowledgment frame upon receiving the data. The coordinator waits for acknowledgment within a timeout range of 1 to 30 seconds, with a preferred timeout of 5 seconds. If no acknowledgment is received, a retransmission is triggered.

[0020] S12. Declaration of Purpose credential verification and digital signature, specifically including: S121. Each teaching node maintains a preset data usage strategy library in its local trusted execution environment. This strategy library is stored in the form of a structured configuration file, with a file size ranging from 1 kilobyte to 10 megabytes, preferably 100 kilobytes. The strategy library contains a mapping rule table between data field identifiers and opcode types, with the number of rules ranging from 10 to 10,000, preferably 500. After receiving the program data operation relationship graph sent by the collaborative computing coordinator, the node performs a step-by-step traversal and comparison of the relationship graph within the trusted execution environment. The traversal order follows the topological sorting of the nodes in the relationship graph, with an upper limit for the traversal depth ranging from 100 to 10,000 levels, preferably 1,000 levels.

[0021] For each record in the relationship graph involving the local privacy data of the teaching node, the node extracts its opcode type and the execution order of that opcode in the control flow, and matches it against the preset rules of the corresponding data field identifier in the local policy library. The matching method uses either exact string matching or regular expression matching; for exact string matching, a complete match is required; for regular expression matching, a similarity threshold is set between 80% and 100%, with a preferred value of 95%. When the similarity reaches the threshold, the opcode type is determined to conform to the local preset data usage policy; if the threshold is not reached, the record is marked as an item awaiting review, triggering a local administrator alarm or an automatic rejection process. After traversal, the node generates a comparison result list, which records the matching status and judgment criteria for each relationship graph record.

[0022] S122. Each teaching node performs a classification operation based on the comparison result list: For opcode types determined to conform to the local preset data usage strategy, the node adds them to the first opcode type whitelist; the number of whitelist entries ranges from 1 to 5000, with a preferred value of 100. For opcode types determined to not conform to the local preset data usage strategy or marked as items pending review, the node adds them to the second opcode type blacklist; the number of blacklist entries ranges from 0 to 5000, with a preferred value of 50. The node labels each opcode type in the whitelist with the corresponding data field identifier, in key-value pair format, with a key length ranging from 1 byte to 256 bytes, with a preferred value of 32 bytes. Simultaneously, the node labels the program position range where each opcode type is allowed to appear in the control flow, represented by the start and end program counter values, with a range span ranging from 1 byte to 10 megabytes, with a preferred value of 1 kilobyte to 100 kilobytes.

[0023] After labeling, the node encapsulates the whitelist, blacklist, data field identifier mapping table, and program location range table in a structured manner to generate a primary-purpose declaration credential. The credential file size ranges from 1 kilobyte to 5 megabytes, with a preferred value of 50 kilobytes. The credential contains a node identifier, timestamp, credential version number, and the aforementioned four types of data tables. The version number ranges from 1.0 to 10.0, with a preferred value of 1.0.

[0024] S123. Each teaching node, within the Trusted Execution Environment (TEE), uses its local private key to digitally sign the first-purpose declaration credential. The private key is stored in the TEE's hardware security module or secure key storage area, with a key length ranging from 128 to 4096 bits, preferably 2048 or 256 bits. The signature algorithm uses a secure hash algorithm or an elliptic curve digital signature algorithm, with a signature output length ranging from 32 to 512 bytes, preferably 64 bytes. The node uses the complete byte sequence of the first-purpose declaration credential as the signature input to calculate the digital signature value. After signing, the node concatenates the original first-purpose declaration credential with the digital signature value to form a signed credential packet, with a packet size ranging from 1 kilobyte to 6 megabytes, preferably 60 kilobytes.

[0025] Nodes send the signed first-purpose declaration credential back to the collaborative computing coordinator via a transport layer secure encrypted channel; the return timeout ranges from 1 second to 60 seconds, with a preferred value of 10 seconds. When the return data packet is transmitted in fragments, the size of a single packet ranges from 1 kilobyte to 64 kilobytes, with a preferred value of 8 kilobytes. After the return is completed, the node retains a local copy, with a retention time ranging from 1 hour to 720 hours, with a preferred value of 72 hours, for subsequent auditing and dispute resolution.

[0026] S13. Generation and distribution of collaborative computing task specifications, specifically including: S131. After receiving the first-purpose declaration credentials from each teaching node, the collaborative computing coordinator verifies the digital signature of each credential individually within a local trusted execution environment or isolated verification sandbox. The verification process calls the public key decryption interface corresponding to the signature algorithm. The public key is extracted from the identity certificate of each teaching node or retrieved from a pre-configured public key directory. The public key length ranges from 128 bits to 4096 bits, with 2048 bits being the preferred value. The coordinator inputs the original byte sequence of the first-purpose declaration credential into the signature verification function and compares it with the accompanying digital signature value. The comparison result is output as a Boolean value. A 100% pass rate for single-credential signature verification is required; if any credential verification fails, the node is marked as an untrusted participant, triggering a retransmission request or removing the node from the participant list.

[0027] After successful verification, the coordinator further verifies the deviation between the timestamp within the credential and the current system time. The tolerance range for deviation is -300s to +300s, with a preferred range of -60s to +60s. Credentials exceeding this tolerance range are considered time-sensitive. For credentials deemed time-sensitive, the collaborative computing coordinator sends a retransmission request to the corresponding teaching node. The maximum number of retransmissions is 1 to 3, with a preferred number of 2. If the retransmission still exceeds the tolerance range, the node is marked as an untrusted participant and removed from the participant list, and its primary purpose declaration credential will not proceed to the subsequent merging process. Simultaneously, the credential version number is verified to be within the supported parsing range. The version number compatibility range is 1.0 to 10.0, with a preferred upper limit of 5.0. After all verifications pass, the credential proceeds to the subsequent merging process.

[0028] S132. The collaborative computing coordinator extracts a whitelist of the first opcode type and a blacklist of the second opcode type from all verified first-purpose declaration credentials for each teaching node. During extraction, the opcode type is categorized according to its string identifier or integer encoding, with a categorization precision requirement of exact matching. For the first opcode type whitelist of all teaching nodes, the coordinator performs an intersection-union operation, retaining only opcode type entries that appear in all node whitelists to obtain a unified opcode type whitelist; the number of entries in this whitelist ranges from 1 to 5000, with a preferred value of 20 to 100. For the second opcode type blacklist of all teaching nodes, the coordinator performs an union-union operation, adding all opcode type entries appearing in any node's blacklist to the union to obtain a unified opcode type blacklist; the number of entries in this blacklist ranges from 0 to 5000, with a preferred value of 10 to 100. During the merging process, if the data field identifiers corresponding to the same opcode type differ in the whitelists or blacklists of different nodes, the coordinator performs normalization processing in the form of a union of the data field identifiers. The upper limit of the number of identifiers after normalization is between 1 and 100, with a preferred value of 5 to 10.

[0029] S133. The collaborative computing coordinator extracts the program position ranges marked by each teaching node for each opcode type from all verified first-purpose declaration credentials. During extraction, the opcode type is used as the index key to read the start and end program counter values ​​marked by each node for that opcode type. For the same opcode type, the coordinator performs an intersection and union operation on the program position ranges marked by all teaching nodes, calculating the maximum value of the start program counter value for each node as the unified starting boundary, i.e.: In the formula: To ensure a consistent initial program counter value, To ensure a consistent end program counter value, The number of teaching nodes participating in the intersection calculation; the minimum value of the end program counter for each node is calculated as the unified end boundary, i.e.: In the formula: For the first The starting program counter value marked at each teaching node. For the first The end program counter value marked at each teaching node; obtain the unified program position range constraint corresponding to the opcode type.

[0030] If a certain opcode type does not have a program location range marked in the first-use declaration document of some nodes, then that node will not participate in the interval intersection calculation for that opcode type. The span value of the unified program location range constraint ranges from 1 byte to 10 megabytes, with a preferred value of 1 kilobyte to 50 kilobytes. When the intersection calculation result is an empty interval, i.e., the unified start boundary is greater than or equal to the unified end boundary, the coordinator marks the opcode type as a location conflict item, triggers an alarm, and suspends the authorization of the opcode type, waiting for renegotiation among the nodes or administrator intervention. After the unified program location range constraints for all opcode types are generated, they are summarized into an interval constraint table.

[0031] S134. The collaborative computing coordinator encapsulates the unified opcode type whitelist, unified opcode type blacklist, unified program location range constraints corresponding to each opcode type, and executable computing programs to generate collaborative computing task specifications. The encapsulation format adopts a structured data container. The container header records the task identifier, generation timestamp, and number of participating nodes. The container body sequentially stores the whitelist table, blacklist table, range constraint table, and executable computing program byte stream. The total size of the container ranges from 1 megabyte to 1 gigabyte, with a preferred value of 20 megabytes to 100 megabytes.

[0032] Once generated, the coordinator sends the collaborative computing task specification to each teaching node via a transport layer secure encrypted channel. The sending method is either broadcast or unicast to each node, with the data packet size per node ranging from 1 kilobyte to 10 megabytes, preferably 500 kilobytes. Each node must return an acknowledgment within a preset time, with a timeout ranging from 1 second to 60 seconds, preferably 10 seconds. Based on the received collaborative computing task specification, each node configures a data access control policy in its local trusted execution environment, loading whitelists, blacklists, and interval constraints into a memory policy cache. The cache refresh cycle ranges from 1 minute to 60 minutes, preferably 5 minutes. Subsequent collaborative computing processes strictly adhere to this specification for data access control.

[0033] S14. Privacy feature decoupling and encapsulation, verifiable credential generation and de-identification encapsulation, specifically including: S141. Each teaching node will input the environmental culture immersion data, activity culture immersion data, and AI interaction immersion data, which have been aligned with the unified educational timeline, into the corresponding modal channel of the multimodal feature extraction model. This multimodal feature extraction model consists of three modality-specific coding subnetworks and one cross-modal feature fusion subnetwork cascaded together. The environmental culture immersion data channel is configured with a one-dimensional time series convolutional coding subnetwork. This subnetwork is composed of one-dimensional convolutional layers, batch normalization layers and max pooling layers stacked sequentially. The number of convolutional kernels ranges from 16 to 256, with a preferred value of 64. The convolutional kernel size ranges from 3 to 15, with a preferred value of 7. The pooling window size ranges from 2 to 4, with a preferred value of 2.

[0034] The activity culture immersion data channel is configured with a category embedding and self-attention encoding sub-network. This sub-network is composed of a category embedding layer, a position encoding layer and a multi-head self-attention layer connected sequentially. The embedding dimension ranges from 16 to 256, with a preferred value of 64. The number of attention heads ranges from 1 to 16, with a preferred value of 4.

[0035] The AI ​​agent interaction immersion data channel is configured with a Transformer encoding subnetwork, which consists of a word embedding layer, a sinusoidal position encoding layer and multiple Transformer encoders stacked together. The number of encoding layers ranges from 2 to 12, with a preferred value of 6. The hidden dimension of the feedforward network ranges from 128 to 2048, with a preferred value of 512.

[0036] The deep semantic feature vectors output by each modality-specific encoding subnetwork have a dimension ranging from 32 to 512, with an optimal value of 128. The cross-modal feature fusion subnetwork employs a weighted feature concatenation layer, which concatenates the deep semantic features of each modality along the channel dimension, and then maps them to a unified feature space via a fully connected layer. Specifically: , In the formula: The intermediate feature vector after weighted concatenation, [ [] is the channel dimension concatenation operator. The deep semantic feature vector output by the environmental and cultural immersion data channel. The deep semantic feature vector output from the data channel for cultural immersion in activities. This provides deep semantic feature vectors output from the data channel that permeates the interaction between artificial intelligence agents. The fusion weights for each modality are given, and satisfy the following conditions: The values ​​range from 0.1 to 0.8, with the default value being equal weight allocation. These fusion weights are automatically learned during the model training phase through end-to-end gradient descent, or are assigned equal weights during the initialization phase. To protect privacy, infiltrate feature vectors. This is the weight matrix of the fully connected layer. This is the bias vector for the fully connected layer. After fusion, a privacy-preserving infiltration feature vector is generated, with a total dimension ranging from 96 to 1536, and an optimal value of 384. The above multimodal feature extraction model is trained end-to-end with supervision using historical campus culture infiltration data and corresponding student digital literacy assessment annotations. The training batch size ranges from 16 to 256, with an optimal value of 64. The number of iterations ranges from 10 to 500, with an optimal value of 100. The optimizer uses an adaptive moment estimation optimizer, with an initial learning rate ranging from 0.0001 to 0.01, and an optimal value of 0.001.

[0037] S142. Before the privacy-preserving infiltration feature vector input path coupling and decoupling model, each teaching node maps the feature vector from the fusion dimension to the coupling residual vector dimension through a total effect mapping layer. The mapping method uses a fully connected layer or linear projection. The input dimension ranges from 96 to 1536, and the output dimension ranges from 16 to 512. The preferred value is a mapping from 384-dimensional input to 128-dimensional output. The total effect mapping layer outputs the total effect vector. Its dimension is consistent with the dimension of the path coupling residual vector. Each teaching node extracts features from the three types of data aligned with the unified educational timeline using a shallow coding subnetwork with the same structure as the modality-specific coding subnetwork corresponding to S141. This yields main effect input features with the same dimension as each modal component of the privacy-preserving infiltration feature vector. These features are then input into the path coupling decoupling model, which consists of a main effect extraction subnetwork, an interaction effect calculation subnetwork, and a residual stripping subnetwork in that order. The main effect extraction subnetwork adopts a three-channel independent multilayer perceptron structure. Each channel corresponds to the environmental culture immersion path, the activity culture immersion path, and the artificial intelligence agent interaction immersion path, respectively. The number of neurons in each perceptron layer ranges from 16 to 512, with an optimal value of 128. The number of layers ranges from 1 to 5, with an optimal value of 3. The activation function adopts rectified linear units or hyperbolic tangent functions, respectively outputting the main effect components of the three types of immersion paths.

[0038] The interaction effect calculation subnetwork employs a bilinear interaction matrix layer. After performing an outer product operation on each pair of the three main effect components, the interaction effect is mapped to a learnable weight matrix. Specifically: In the formula: For path With path The interaction effect size between them For path With path The main effect components, The outer product matrix of the main effect components. For matrix vectorization operators, For paths and paths The corresponding learnable weight matrix, For path With path The corresponding bias vector. The rank of the interaction matrix ranges from 1 to 64, with an optimal value of 8. The dimension of the output interaction effect size is consistent with the dimension of the main effect components.

[0039] The residual stripping subnetwork employs a vector subtraction aggregation layer. It subtracts each main effect component and interaction effect from the total effect vector output by the total effect mapping layer for the three types of data, extracting the residual nonlinear coupling term as the path coupling residual vector. Specifically: In the formula: This is the path coupling residual vector. The total effect vector output by the total effect mapping layer. For the set of infiltration paths, For path The main effect components, For path With path The interaction effect size. The dimension of the coupled residual vector ranges from 16 to 512, with a preferred value of 128. The element range is from negative infinity to positive infinity, with typical values ​​ranging from -1.0 to +1.0. After L2 norm normalization, the vector magnitude ranges from 0.5 to 2.0, with a preferred value of 1.0. In the formula: This is the path coupling residual vector after L2 norm normalization. Let L2 be the norm of the vector. The above path coupling and decoupling model is trained end-to-end with supervision using historical campus culture immersion data and corresponding student digital literacy assessment labels. The training batch size ranges from 16 to 256, with an optimal value of 64. The number of iterations ranges from 10 to 500, with an optimal value of 100. The optimizer is an adaptive moment estimation optimizer, and the initial learning rate ranges from 0.0001 to 0.01, with an optimal value of 0.001.

[0040] S143. Each teaching node generates a local proof within the trusted execution environment based on the generation process of privacy-preserving infiltrated feature vectors and coupled residual vectors. The generation process of the local proof records the model version number, inference timestamp, input data hash value, and output vector hash value called during feature extraction and coupling decoupling. The model version number ranges from 1.0 to 100.0, with a preferred value of 1.0. The hash algorithm adopts a series of secure hash algorithms, and the output length ranges from 16 bytes to 64 bytes, with a preferred value of 32 bytes.

[0041] Nodes utilize zero-knowledge proof algorithms to transform local proofs into verifiable credentials. The zero-knowledge proof algorithm employs either zk-SNARKs or zk-STARKs schemes. The proof generation time ranges from 0.1s to 600s, with a preferred value of 10s. The proof size ranges from 100 bytes to 100 kilobytes, with a preferred value of 1 kilobyte. The verifiable credential contains public verification parameters and zero-knowledge proof ciphertext. Verification parameters include a node identifier, task identifier, generation timestamp, and output vector commitment value. Nodes store the verifiable credential in a protected memory region within a trusted execution environment. This region's size ranges from 1 megabyte to 1 gigabyte, with a preferred value of 10 megabytes. The credential's validity period ranges from 1 hour to 720 hours, with a preferred value of 72 hours. Expired credentials automatically expire and trigger regeneration.

[0042] S144. At each teaching node, differential privacy noise perturbations are applied to the privacy-preserving infiltrated feature vector and the coupled residual vector, respectively. The differential privacy mechanism adopts a Gaussian mechanism, specifically... , In the formula: The perturbation vector is the first dimensional elements, For the original vector of the th dimensional elements, For the injection of the first Gaussian noise of dimension 1 The standard deviation of Gaussian noise. To protect privacy, infiltrate feature vectors or couple residual vectors Sensitivity, with a value ranging from 0.01 to 10.0, and a default value of 1.0; The privacy budget parameter ranges from 0.1 to 10.0, with a preferred value of 1.0. This is the differential privacy relaxation parameter, with a value range of 10. −12 Up to 10 −5 The default value is 10. −6 Or 10 −9 The noise scale is based on the privacy-preserving infiltration feature vector or coupled residual vector. Sensitivity calculate.

[0043] The node associates and binds the perturbed feature vector and coupled residual vector with the data usage tags corresponding to the collaborative computing task specifications. The binding method uses a structured dictionary or key-value mapping table. The data usage tag records the mapping association between each dimension of the privacy-preserving infiltrated feature vector and coupled residual vector and the target data field identifier. The number of mapping table entries ranges from 10 to 10,000, with a preferred value of 500. The data usage tag also records the corresponding Unicode type whitelist, Unicode type blacklist, and program location range constraints. The tag file size ranges from 1 kilobyte to 1 megabyte, with a preferred value of 50 kilobytes. After completing the association binding, the node serializes and encapsulates the binding result to generate a de-identified encapsulated data packet. The packet size ranges from 1 kilobyte to 10 megabytes, with a preferred value of 100 kilobytes, thus completing the de-identification encapsulation.

[0044] S2. Collaborative nodes combine privacy-preserving infiltrated feature vectors to construct a policy heterogeneity subspace; counterfactual causal inference calculates cross-node net attribution contribution and policy heterogeneity calibration factors; the distributed consensus network blocks unauthorized access and generates usage execution audit records; specifically including: S21. Cross-domain alignment and strategy heterogeneity subspace construction, specifically including: S211. Each teaching node encapsulates the privacy-preserving infiltration feature vector, coupled residual vector, and verifiable credentials into a data distribution package within its local trusted execution environment. The encapsulation format adopts a structured binary container. The container header records the node identifier, task identifier, timestamp, and data hash check value. The container body sequentially stores the privacy-preserving infiltration feature vector, coupled residual vector, and verifiable credential ciphertext. The size of the distribution package ranges from 10 kilobytes to 100 megabytes, with 1 megabyte being the preferred value.

[0045] Nodes transmit published packets to the distributed consensus network via a trusted data delivery protocol. This protocol includes at least three stages: encrypted data transmission, sender digital signature, and receiver acknowledgment. Encrypted transmission uses Advanced Encryption Standard (AES) or Transport Layer Security (TLS) protocols, with a key length ranging from 128 to 256 bits, preferably 256 bits. The sender's digital signature uses an elliptic curve digital signature algorithm, with a signature length ranging from 64 to 128 bytes, preferably 64 bytes. The receiver acknowledgment includes a reception timestamp, a node status code, and a data integrity check value; the acknowledgment timeout ranges from 1 to 60 seconds, preferably 10 seconds. If no acknowledgment is received, the node triggers automatic retransmission, with a maximum number of retransmissions ranging from 1 to 10, preferably 3.

[0046] S212. After receiving data distribution packets from each teaching node within the distributed consensus network, the collaborative node performs zero-knowledge proof verification on the verifiable credentials in an isolated verification environment. The verification process calls the zero-knowledge proof verification algorithm, inputting public verification parameters and zero-knowledge proof ciphertext, and outputting a Boolean value as the verification result. The verification content includes whether the generation process of the privacy-preserving infiltrated feature vector and the coupled residual vector is completed within a trusted execution environment, whether the generation timestamp is within a valid range, and whether the output vector commitment value is consistent with the data in the distribution packet. The verification time ranges from 0.1s to 60s, with a preferred value of 5s. A 100% pass rate is required for a single verification; that is, if any verification item fails, the teaching node's data is deemed untrustworthy, the node is marked as an abnormal node, its data participation in subsequent calculations is suspended, and an alarm notification is triggered. The upper limit for the percentage of abnormal nodes ranges from 0% to 50%, with a preferred value of 20%. When the percentage of abnormal nodes exceeds the upper limit, the collaborative node determines that the current collaborative computation task is not feasible and returns a task termination instruction to all teaching nodes.

[0047] S213. After successful verification, the collaborative node extracts time dimension information from the privacy-preserving immersion feature vectors of each teaching node. This time dimension information includes the sampling timestamp, the cultural immersion time window index, and the educational timeline offset. The collaborative node constructs a unified time coordinate system based on the educational timeline, with a resolution ranging from 1 minute to 1 hour, ideally 15 minutes. For the time dimension information of each teaching node, the collaborative node performs a time-series alignment operation, using nearest-neighbor interpolation or linear time interpolation. The interpolation accuracy error tolerance range is from 0 seconds to 300 seconds, ideally 60 seconds.

[0048] The collaborative nodes, combining the coupled residual vectors of each teaching node, adjust the phase shift of the feature data of each node in a unified time coordinate system based on the temporal coupling relationship of the education time axis. The temporal coupling relationship is determined by calculating the cross-correlation function of the coupled residual vectors of each node, specifically: where: For the first The teaching node and the first The coupled residual vector of each teaching node in time delay Normalized cross-correlation coefficients under the following conditions For the first Each teaching node at time The coupled residual vector, Let be the mean vector of the coupled residual vector of the i-th teaching node during the observation period. This represents the total number of time steps within the observation period. This is the time delay offset, with a value ranging from 0 to... In time delay offset Within the range of values, the cross-correlation function is calculated by iterating through the data with a preset step size, and the maximum value of the function is taken as the cross-correlation peak. The cross-correlation peak threshold ranges from 0.3 to 0.9, with a preferred value of 0.7. Node pairs with peak values ​​below the threshold are marked as weakly coupled and a conservative alignment strategy is adopted. After alignment, the privacy-preserving infiltration feature vector and coupling residual vector of each teaching node are mapped to a unified time coordinate system to form a cross-domain alignment feature matrix, where the number of rows in the matrix represents the number of time steps and the number of columns represents the feature dimension.

[0049] S214. In the feature space where cross-domain alignment is completed, the collaborative nodes perform sub-vector segmentation on the privacy-preserving infiltrated feature vectors according to the dimensionality index before modal fusion. The segmentation method is consistent with the splicing order of the cross-modal feature fusion sub-network in S141, i.e., the previous... The corresponding environmental and cultural immersion subspace, in the middle The corresponding activity cultural immersion subspace, after The dimensions correspond to the AI ​​agent interaction immersion subspace, where the value of each dimension ranges from 16 to 512, with an optimal value of 128. The feature distribution differences among teaching nodes in the environmental culture immersion dimension, activity culture immersion dimension, and AI agent interaction immersion dimension are calculated. The distribution difference is calculated using Euclidean distance, with the distance metric ranging from 0 to positive infinity, and typical values ​​ranging from 0.1 to 10.0. Specifically: In the formula: For the first The teaching node and the first Each teaching node in the immersion dimension Euclidean distance on, For immersive dimension indexing, For the first Each teaching node in the immersion dimension The value of the i-th feature dimension, For immersion dimension The total number of feature dimensions. The collaborative node compares the feature distribution differences in each dimension with a preset threshold, which ranges from 0.1 to 5.0, with a preferred value of 2.0.

[0050] For dimensions where the difference in feature distribution exceeds a preset threshold, the collaborative node extracts the feature offset vector of the corresponding teaching node in that dimension. The feature offset vector dimension ranges from 16 to 512, with a preferred value of 128. The feature offset vector is calculated by subtracting the arithmetic mean of the subvectors of all teaching nodes in that dimension from the subvector of the teaching node in that immersion dimension. In the formula: For the first Each teaching node in the immersion dimension Feature offset vector on, Let it be its subvector. This represents the total number of teaching nodes. Collaborative nodes aggregate feature offset vectors exceeding a threshold across all dimensions. Aggregation methods include principal component analysis for dimensionality reduction or cluster center concatenation. After dimensionality reduction, the subspace dimension ranges from 2 to 128, with an optimal value of 16. After aggregation, collaborative nodes construct a policy heterogeneity subspace. Each basis vector in this subspace corresponds to a policy difference pattern. The number of basis vectors in the subspace ranges from 2 to 64, with an optimal value of 8. This completes the construction of the policy heterogeneity subspace.

[0051] S22. Counterfactual causal attribution and calibration factor calculation, specifically including: S221. Within the heterogeneous strategy subspace, the collaborative nodes construct a virtual control strategy for each teaching node: The collaborative nodes first read the strategy parameter combination corresponding to the actual campus culture immersion strategy adopted by each teaching node. This strategy parameter combination includes the environmental culture immersion intensity coefficient, the activity culture immersion frequency parameter, and the AI ​​agent interaction response threshold. The number of parameters ranges from 3 to 30, with a preferred value of 6. The preset control strategy construction rule adopts either the single-path perturbation method or the temporal offset method. The single-path perturbation method multiplies the strategy intensity of the target immersion path by the perturbation coefficient, i.e.: In the formula: For the first The immersion path of each teaching node objective The strength of the perturbed virtual control strategy For the first The actual target immersion path adopted at each teaching node Strategy strength, The disturbance coefficient ranges from 0.5 to 1.5, with a preferred value of 0.8 or 1.2. The time-series offset method shifts the strategy initiation time of the target infiltration path by a preset duration, i.e.: In the formula: For the first The immersion path of each teaching node objective Startup time of the offset virtual control strategy For the first The actual target immersion path adopted at each teaching node Strategy startup time, This is the time offset, with a value ranging from -7 days to +7 days, and the preferred value being -3 days or +3 days.

[0052] The virtual control strategy only changes the strategy strength or timing of the target infiltration path, while keeping the strategy parameters of other paths unchanged. Each teaching node independently generates a virtual control strategy for each target infiltration path, perturbing only a single path at a time while keeping the actual strategy parameters of the other paths unchanged. For a single target infiltration path, the number of virtual control strategies generated ranges from 1 to 10, with a preferred value of 3. That is, multiple virtual control strategies are generated for the same target infiltration path using different perturbation coefficients or different offset durations for robustness verification. After generation, the collaborative node pairs and stores the virtual control strategies with the actual strategies as strategy pairs for subsequent attribution evaluation.

[0053] S222. The collaborative node invokes the attribution evaluation function to predict students' digital literacy growth for each policy pair. This attribution evaluation function is composed of a policy parameter embedding subnetwork, a feature spatiotemporal coding subnetwork, a policy-feature cross-attention subnetwork, and a regression prediction subnetwork, all cascaded sequentially. The policy parameter embedding subnetwork employs a multilayer perceptron structure, mapping policy parameter combinations to a dense vector space. The number of hidden layers ranges from 1 to 4, with a preferred value of 2 layers. The number of neurons per layer ranges from 16 to 256, with a preferred value of 64. The output embedding vector dimension ranges from 32 to 512, with a preferred value of 128. The feature spatiotemporal encoding subnetwork uses a stacked structure of one-dimensional time-series convolutional layers and bidirectional long short-term memory layers to temporally encode privacy-preserving infiltrated feature vectors. The number of convolutional kernels ranges from 16 to 128, with a preferred value of 32. The long short-term memory hidden dimension ranges from 32 to 256, with a preferred value of 64. The output encoding vector dimension ranges from 64 to 1024, with a preferred value of 256. The policy-feature cross-attention subnetwork employs a multi-head cross-attention mechanism, using the policy embedding vector as the query sequence and the feature encoding vector as the key-value sequence. It models the nonlinear interaction between policy parameters and features. The number of attention heads ranges from 1 to 8, with a preferred value of 4. The output interaction vector dimension ranges from 64 to 512, with a preferred value of 128. The regression prediction subnetwork uses a fully connected layer structure, mapping the interaction vector to predicted student digital literacy growth values. The output layer has 1 neuron, and the output value range is 0 to 1 (normalized growth index), with a preferred value distribution between 0.4 and 0.8. The attribution evaluation function is trained end-to-end using historical campus culture immersion data, actual policy parameters, and corresponding student digital literacy evaluation labels. The training batch size ranges from 16 to 256, with a preferred value of 64. The number of iterations ranges from 10 to 500, with a preferred value of 100. The optimizer uses an adaptive moment estimation optimizer, with an initial learning rate ranging from 0.0001 to 0.01, with a preferred value of 0.001.

[0054] The collaborative nodes input the input data corresponding to the actual strategy into the attribution evaluation function and output the first predicted value; they input the input data corresponding to the virtual control strategy into the same function and output the second predicted value; the difference between the two is calculated as the node attribution contribution of each teaching node on the corresponding target immersion path, i.e.: In the formula: For the first Each teaching node in the target immersion path Node attribution contribution For the attribution evaluation function on the first Target Immersion Path under the Actual Strategies of Each Teaching Node The first predicted value output, For the attribution evaluation function on the first Target Immersion Path under Virtual Comparison Strategy at Each Teaching Node The second predicted value output. When the node's attribution contribution... When the absolute value exceeds 0.2, the node is marked as a strongly influential node.

[0055] S223. Collaborative nodes perform weighted aggregation of the attribution contributions of all nodes along the same target infiltration path. Before aggregation, collaborative nodes obtain the local sample size and strategy execution duration from each teaching node. The local sample size ranges from 10 to 10,000, with an optimal value of 500. The strategy execution duration ranges from 7 to 365 days, with an optimal value of 90 days. Collaborative nodes calculate normalized weights. The weight formula is the product of the local sample size and the strategy execution duration divided by the sum of the products of all nodes. That is, the weight is proportional to both the sample size and the duration. Specifically: In the formula: For the first Normalized weights of each teaching node, For the first Local sample size of each teaching node For the first The strategy execution time for each teaching node. The total number of teaching nodes participating in the weighted aggregation; the weight value ranges from 0.001 to 1.0, with the preferred value ranging from 0.01 to 0.3, which is suitable for typical scenarios with 5 to 10 nodes.

[0056] The collaborative nodes multiply the node attribution contribution of each teaching node by its corresponding normalized weight and then sum them to obtain the cross-node net attribution contribution of the target immersion path, i.e.: In the formula: Target infiltration path Cross-node net attribution contribution For the first Normalized weights of each teaching node, For the first Each teaching node in the target immersion path The node-based attribution contribution. When the net attribution contribution across nodes... When the absolute value is less than 0.08, the target immersion path is determined to have no significant contribution to the overall growth of students' digital literacy, and is marked as a low contribution path with a reduced priority in subsequent resource allocation.

[0057] S224. Collaborative Node Calculation Strategy Heterogeneity Calibration Factor: Collaborative nodes first statistically analyze the node attribution contribution of each teaching node on different target immersion paths, and calculate its dispersion. The dispersion is measured using the standard deviation, i.e.: In the formula: Target infiltration path The standard deviation of the attribution contribution of each teaching node ranges from 0 to 2, with the preferred value ranging from 0.05 to 0.5. Target infiltration path The arithmetic mean of the attribution contributions of all teaching nodes, i.e. .

[0058] The collaborative nodes divide the degree of dispersion by the baseline value of the strategy differences among the teaching nodes to obtain the influence weight of the differences in campus culture immersion strategies among the teaching nodes on the attribution results, that is: In the formula: Target infiltration path The heterogeneity of strategies affects the weights. As the baseline value for the strategic differences between each teaching node, the arithmetic mean of the standard deviations of all target immersion paths is taken, i.e. , The total number of target infiltration paths; when the influence weight When the value is greater than 2.0, it is determined that policy heterogeneity significantly interferes with the attribution results. The collaborative nodes take the arithmetic mean of the influence weights of each target infiltration path to generate a policy heterogeneity calibration factor, i.e.: In the formula: This is the calibration factor for strategy heterogeneity. When the calibration factor... When the deviation from 1.0 exceeds 0.5, the cross-node policy negotiation mechanism is triggered, requiring nodes with significant differences to readjust their local policy parameters.

[0059] S23. Usage-constrained access control and audit log generation, specifically including: S231. The distributed consensus network deploys a real-time access monitoring proxy throughout the attribution computation process. This proxy intercepts every feature dimension access request initiated by the cooperating nodes using a polling or event-triggered method. The monitoring frequency ranges from once per millisecond to once per second, with a preferred value of once every 10 milliseconds. After interception, the proxy parses the request data packet header, extracting the target dimension identifier and the request opcode type. The target dimension identifier is in string or integer index form, with a length ranging from 1 byte to 256 bytes, with a preferred value of 32 bytes. The request opcode types include five basic types: read opcode, write opcode, aggregation opcode, transmission opcode, and delete opcode.

[0060] After extraction, the proxy encapsulates the target dimension identifier and request opcode type into an access audit record. The record format uses structured key-value pairs, and the size of a single record ranges from 64 bytes to 1 kilobyte, with a preferred value of 256 bytes. The proxy stores the access audit record in a first-in-first-out audit queue, with a queue depth ranging from 100 to 10,000 records, with a preferred value of 1,000 records. When the queue depth exceeds the upper limit, the earliest enqueued record is automatically discarded, and a queue overflow alarm is triggered.

[0061] S232. The distributed consensus network uses the target dimension identifier as the index key to query the data usage tag attached to that dimension in a distributed hash table or in-memory index structure. The index query response time ranges from 0.1ms to 100ms, with a preferred value of 5ms. Upon a successful query, the network parses three core constraints from the data usage tag: The first is a whitelist of Unicode opcode types, containing the set of allowed opcode types for that dimension, with a count ranging from 1 to 100, preferably 10. The second is a blacklist of Unicode opcode types, containing the set of prohibited opcode types for that dimension, with a count ranging from 0 to 100, preferably 5. The third is a program location range constraint, containing the start and end program counter values ​​allowed for access to that dimension, with a range span ranging from 1 byte to 10 megabytes, preferably 1 kilobyte. After parsing, the network compares these three constraints with the current request opcode type and request initiation location.

[0062] S233. The distributed consensus network performs a dual-condition judgment: The first judgment checks whether the request opcode type belongs to the unified opcode type whitelist. If it does, the judgment is passed; otherwise, it is judged as a violation. The second judgment checks whether the request opcode type does not belong to the unified opcode type blacklist. If it does not belong, the judgment is passed; otherwise, it is judged as a violation. When both judgments pass, the access request for that feature dimension is allowed to continue. If either judgment fails, the network immediately determines that the access request for that feature dimension exceeds the scope declared in the collaborative computing task specification and triggers a blocking instruction. The blocking instruction directly intercepts the data stream through the underlying access control interface. The blocking response time ranges from 0.01ms to 10ms, with a preferred value of 0.1ms. After successful blocking, the collaborative node's access operation on the target feature dimension is forcibly terminated, and the access result returns a null value or an abnormal status code. The status code ranges from 400 to 599, with a preferred value of 403. If blocking fails, a security alarm is triggered and all data access permissions for the collaborative node are suspended.

[0063] S234. After completing the blocking operation, the distributed consensus network immediately collects audit metadata, including the identifier of the teaching node that initiated the access request for that feature dimension, the identifier of the accessed target feature dimension, the type of the blocked request opcode, and the precise blocking timestamp. The teaching node identifier uses a globally unique identifier format, with a length ranging from 16 bytes to 64 bytes, preferably 32 bytes. The blocking timestamp uses Coordinated Universal Time (UTC) format with millisecond precision. The network associates these four metadata items with the identifier of this collaborative computing task to generate a purpose execution audit record. The size of a single audit record ranges from 128 bytes to 2 kilobytes, preferably 512 bytes. The network writes the purpose execution audit record to the audit log storage area of ​​the distributed consensus network. This area uses an append-only write mode, with a storage capacity ranging from 1 gigabyte to 100 gigabytes, preferably 10 gigabytes. After writing, the network synchronizes the purpose execution audit record to all teaching nodes via a broadcast protocol, with a synchronization delay ranging from 10ms to 30s, preferably 1s. Each node returns an acknowledgment upon receiving the record.

[0064] S3. At each teaching node, detect the critical transition point for students' digital literacy development, calculate the policy deviation between the policy prediction output and the cross-node net attribution contribution, generate an optimized gradient update model by combining the policy heterogeneity calibration factor, and backfeed the parameter summary and transition threshold to the distributed consensus network; specifically including: S31. Critical transition detection and strategy deviation calculation, specifically including: S311. Each teaching node invokes a pre-deployed campus culture immersion strategy model within its local trusted execution environment. This model consists of a sequentially cascaded feature temporal coding subnetwork, a coupled residual fusion subnetwork, and a literacy development prediction subnetwork. The feature temporal coding subnetwork employs gated recurrent units or long short-term memory networks, using the verified privacy-preserving infiltrated feature vector as input to extract temporal dependencies. The hidden layer dimension ranges from 32 to 512, with an optimal value of 128, and the number of layers ranges from 1 to 4, with an optimal value of 2. The coupled residual fusion subnetwork employs gated linear units or attention-gated mechanisms to nonlinearly fuse the temporal hidden state output by the feature temporal coding subnetwork with the coupled residual vector. Specifically: , In the formula: For a moment The fused hidden state output by the coupled residual fusion subnetwork. For the feature temporal coding subnetwork at time... Output of the timing hidden state, For a moment The coupled residual vector, For a moment The gating weight ranges from 0 to 1, with a preferred value of 0.5. It is the Sigmoid activation function. For the gated weight matrix, For gating bias terms, This is a vector concatenation operation. The literacy growth prediction subnetwork adopts a fully connected layer structure. The output layer is followed by a Sigmoid activation function, which maps the fused features to the time-series predicted values ​​of students' digital literacy growth and compresses them to the range of 0 to 1. The number of neurons in the output layer ranges from 1 to 10, with an optimal value of 1. The output value range is 0 to 1 (normalized growth index). The above campus culture immersion strategy model uses historical campus culture immersion data and corresponding student digital literacy assessment labels for end-to-end supervised training. The training batch size ranges from 16 to 256, with an optimal value of 64. The number of iterations ranges from 10 to 500, with an optimal value of 100. The optimizer adopts an adaptive moment estimation optimizer, and the initial learning rate ranges from 0.0001 to 0.01, with an optimal value of 0.001.

[0065] The node performs audit logging based on its purpose, counts the number of blocked access requests for specific feature dimensions, and calculates a collaborative computing trust score. The scoring formula is as follows: In the formula: To collaboratively compute credibility scores, This is the arithmetic mean of the number of feature dimension access requests blocked for each teaching node during this collaborative computing process. As a reference blocking threshold, it is defined as 1.5 times the average number of blocked requests in historical collaborative computing tasks at each teaching node, with a value ranging from 10 to 10000, and a preferred value of 100; when When, the score is 0; when At that time, the score is set to 1. The node uses this score to apply a confidence-weighted average to the cross-node net attribution contribution, obtaining the weighted cross-node net attribution contribution, i.e.: In the formula: The weighted net attribution contribution across nodes. For cross-node net attribution contribution, To score the credibility of collaborative computing.

[0066] The node performs usability verification on the privacy-preserving infiltration feature vector based on this score, generating a verified vector, specifically: , In the formula: To protect privacy, the proportion of missing dimensions in the infiltration feature vector. This represents the number of feature dimensions missing due to cross-node blocking. The total dimension of the privacy-preserving infiltrated feature vector. This is the verified privacy-preserving infiltration feature vector. To protect privacy, infiltrate feature vectors. For the validity mask vector, its generation rule is: for the If the data for that dimension is not missing, then If the data for this dimension is missing and the proportion of missing dimensions is high... ,but ;like If the privacy protection infiltration feature vector is determined to be unusable, the current transition detection process will be terminated and a data completion request will be triggered. The padding vector is generated by zero-value padding or neighborhood mean interpolation. This is element-wise multiplication.

[0067] The node inputs the verified vector into the feature temporal coding subnetwork and the coupled residual vector into the coupled residual fusion subnetwork. The literacy development prediction subnetwork outputs a temporal prediction sequence, which is used to detect the critical transition point in students' digital literacy development, i.e.: In the formula: For a moment Predicted values ​​of students' digital literacy growth The output weight matrix of the literacy development prediction subnetwork. This is the output bias term of the literacy growth prediction subnetwork.

[0068] S312. Each teaching node uses the education timeline as the time series benchmark, and extracts the verified vector in segments according to a preset sliding time window. The sliding time window length ranges from 5 to 50 time steps, with a preferred value of 20 time steps. The window sliding step size ranges from 1 to the window length times, and does not exceed half the window length, with a preferred value of half the window length (rounded down). Each node calculates the temporal rate of change of students' digital literacy growth within each time window. The rate of change is calculated by dividing the first difference of the predicted values ​​of adjacent time steps by the time interval, i.e. In the formula: For a moment The rate of change in students' digital literacy over time. A model for campus culture immersion strategies at all times Output of student digital literacy growth prediction values This represents the time interval between adjacent time steps. A node takes the arithmetic mean of the temporal change rates of all adjacent time steps within the time window as the representative temporal change rate of that window, or takes the maximum temporal change rate within the window as the representative temporal change rate of that window.

[0069] The node calculates the nonlinear interaction intensity coefficient based on the L2 norm of the coupling residual vector, and reads the path coupling residual vector normalized by the L2 norm. Calculate its L2 norm as the intensity coefficient of the nonlinear interaction effect. If this vector has been strictly normalized, then... Taking 1.0 corrects for degradation to an identity mapping; if an unnormalized or approximately normalized vector is used, then... The value ranges from 0.5 to 1.5, with a preferred value of 1.0. A scalar-weighted correction is applied to the time-series rate of change; the corrected rate of change is the product of the original rate of change and the intensity coefficient, i.e.: In the formula: This is the corrected time series rate of change. Let L2 norm be the coupled residual vector. For a moment The time-series change rate of students' digital literacy development. The node dynamically compares the corrected time-series change rate with a locally dynamically adjusted transition threshold. The initial value of the transition threshold ranges from 0.01 to 0.5, with a preferred value of 0.1. The threshold is adaptively adjusted based on historical transition frequencies, with an adjustment step size ranging from 0.001 to 0.01. The adaptive adjustment rule is: if continuous... If no transition point is detected within a time window, the transition threshold is reduced by one step. If in a continuous Detected more than within a time window If there are any invalid transition points (where the criterion for an invalid transition point is that the absolute value of the predicted change in students' digital literacy growth within that time window is less than 0.05), then the transition threshold is increased by one step. ;in The value ranges from 5 to 20, with a preferred value of 10. The value ranges from 3 to 10, with 5 being the preferred value. The value ranges from 2 to 5, with a preferred value of 3; adjust the step size. The value ranges from 0.001 to 0.01, with a preferred value of 0.005; after threshold adjustment, it is limited to the range of 0.01 to 0.5. When the corrected temporal change rate exceeds the transition threshold in multiple consecutive time windows, the number of consecutive windows ranges from 2 to 10, with a preferred value of 3. The node determines the starting time of the first window that meets the consecutive exceedance condition as the critical transition point for the growth of students' digital literacy.

[0070] S313. Each teaching node obtains the first strategy prediction output generated by the campus culture immersion strategy model based on local feature input within the most recent time window before the critical transition point, and the second strategy prediction output generated by the campus culture immersion strategy model based on local feature input within the most recent time window after the critical transition point; both the first and second strategy prediction outputs are taken as the predicted value of students' digital literacy growth at the last moment within the corresponding time window. Both the first and second strategy prediction outputs are numerical results obtained by the campus culture immersion strategy model through forward inference on the locally validated vectors, with data type being 32-bit floating-point numbers and a value range of 0 to 1 (normalized growth index).

[0071] The node calculates the difference between the predicted output of the second strategy and the predicted output of the first strategy to obtain the predicted change in students' digital literacy growth, i.e.: In the formula: To predict changes in students' digital literacy growth. This is the second policy prediction output within the most recent time window after the critical transition point. This is the first policy prediction output within the most recent time window before the critical transition point. When the change... When the absolute value is less than 0.05, the critical transition point is determined to be noise-triggered, marked as an invalid transition, and subsequent deviation calculations are skipped. When the absolute value of the change is greater than 0.5, an anomaly alarm is triggered and the model diagnostic process is started.

[0072] S314. At each teaching node, the policy bias is calculated by subtracting the predicted change in students' digital literacy growth from the weighted cross-node net attribution contribution. In the formula: This is a strategy bias. To predict changes in students' digital literacy growth. This represents the weighted net attribution contribution across nodes.

[0073] The node stores the policy bias in a temporary result cache within its local trusted execution environment. The cache size ranges from 1 megabyte to 100 megabytes, with a preferred value of 10 megabytes. When the absolute value of the policy bias exceeds a preset alarm threshold (ranging from 0.1 to 1.0, with a preferred value of 0.1), the node triggers an excessive policy bias alarm, indicating a significant deviation between the local model and the cross-node attribution results, requiring the initiation of a model calibration process. At this point, the node has completed the calculation of the policy bias between the policy prediction output of the campus culture immersion strategy model before and after the critical transition point and the net attribution contribution across nodes.

[0074] S32. Targeted optimization gradient generation and remote trusted backhaul, specifically including: S321. Each teaching node reads the policy bias from the temporary cache of its local trusted execution environment and reads the policy heterogeneity calibration factor. The node calculates the reciprocal of the policy heterogeneity calibration factor, i.e., 1 divided by the factor. When the factor equals 1.0, the reciprocal is still 1.0; when the factor is greater than 1.0, the reciprocal is less than 1.0; and when the factor is less than 1.0, the reciprocal is greater than 1.0. The node multiplies the policy bias by this reciprocal to obtain the bias value after preliminary calibration. The node further reads the collaborative computing trustworthiness score, multiplies the bias value after preliminary calibration by the collaborative computing trustworthiness score, and obtains the final calibrated policy bias, specifically: In the formula: represents the strategy bias. As a policy heterogeneity calibration factor, To collaboratively compute credibility scores, The final calibration policy bias eliminates interference from cross-node policy differences and corrects the credibility loss caused by data access blockage during collaborative computation. Nodes store the final calibration policy bias in a local optimization cache, with a cache size ranging from 1 megabyte to 100 megabytes, preferably 10 megabytes, providing an optimization target for subsequent gradient generation.

[0075] S322. Each teaching node aims to minimize the final calibration strategy deviation. It utilizes the backpropagation mechanism of the campus culture immersion strategy model. Backpropagation begins at the output layer of the literacy growth prediction sub-network and calculates the gradient of the loss function with respect to each model parameter layer by layer. Specifically: where g is the original gradient vector of the loss function with respect to the model parameters. This is the set of trainable parameters for a campus culture immersion strategy model. The loss function takes the calibration strategy bias as input and adopts the mean squared error form. Or mean absolute error form The loss value ranges from 0 to 4, with the preferred value ranging from 0 to 0.5; the learning rate ranges from 0.0001 to 0.1, with the preferred value being 0.001; the gradient calculation covers all trainable parameters of the feature temporal coding subnetwork, the coupled residual fusion subnetwork, and the literacy growth prediction subnetwork, including the weight matrix and the bias vector.

[0076] The node performs L2 norm clipping on the calculated gradient direction, specifically: In the formula: represents the directional optimization gradient after L2 norm clipping. The gradient clipping threshold ranges from 0.1 to 10.0, with a preferred value of 1.0 to prevent gradient explosion. The L2 norm of the original gradient vector is given. The clipped gradient is the directed optimization gradient, whose direction points towards the parameter adjustment direction that makes the policy prediction output of the literacy growth prediction subnetwork approach the net attribution contribution across nodes. The dimension of the directed optimization gradient is consistent with the total number of model parameters, ranging from -1.0 to +1.0, with the preferred value ranging from -0.01 to +0.01. Nodes store the directed optimization gradient in a local gradient cache, with the cache size ranging from 1 megabyte to 100 megabytes, and the preferred value being 10 megabytes.

[0077] S323. Each teaching node performs gradient descent updates on the current parameters of the campus culture immersion strategy model according to the directed optimization gradient. The update formula is the current parameters minus the product of the learning rate and the directed optimization gradient, i.e.: In the formula: For the updated policy model parameters, The current parameters for the campus culture immersion strategy model are: For learning rate, To optimize gradients in a targeted manner, the update range covers all weight matrices and bias vectors of the feature temporal coding subnetwork, the coupled residual fusion subnetwork, and the literacy growth prediction subnetwork. Nodes perform single-step updates or multi-step iterative updates, with the number of iterations ranging from 1 to 100, and the preferred value being 10.

[0078] After the update, the node obtains the updated policy model parameters. The node then performs dimensionality reduction on the updated parameters, using principal component analysis or random projection to map the high-dimensional parameters to a low-dimensional summary space. The dimensionality after reduction ranges from 16 to 1024, with a preferred value of 128. The node concatenates the dimensionality-reduced parameter vectors into a parameter summary, with a summary length ranging from 64 bytes to 8 kilobytes, with a preferred value of 512 bytes. Simultaneously, the node merges the transition thresholds corresponding to critical transition points as independent parameters with the parameter summary to construct the data packet to be transmitted back. The total size of the data packet ranges from 128 bytes to 10 kilobytes, with a preferred value of 1 kilobyte. The node stores the data packet to be transmitted back in a remote proof cache, with a cache size ranging from 1 megabyte to 100 megabytes, with a preferred value of 10 megabytes.

[0079] S324. Each teaching node invokes the remote proof module of its local trusted execution environment, using the data packet to be transmitted as the additional payload input to the remote proof report. The remote proof module reads the unique endorsement key from the trusted platform module or trusted execution environment hardware. The key length ranges from 2048 bits to 4096 bits, with 2048 bits being the preferred value. The module generates a remote proof report containing the current value of the platform configuration register, the hash digest of the data packet to be transmitted, and a timestamp. The hash algorithm uses a secure hash algorithm, with a digest length ranging from 32 bytes to 64 bytes, with 32 bytes being the preferred value. Specifically: In the formula: This is the hash digest of the data packet to be transmitted back. The data packet to be transmitted is constructed by combining the parameter digest and the transition threshold. This is a secure hash algorithm. The module uses an endorsement key to hardware-sign the remote proof report, generating a hardware-signed remote proof report; the signature length ranges from 64 bytes to 512 bytes, with a preferred value of 256 bytes.

[0080] Nodes transmit hardware-signed remote proof reports back to the distributed consensus network via a transport layer secure encrypted channel. The transmission timeout ranges from 1 to 60 seconds, with a preferred value of 10 seconds. Upon receiving the report, other teaching nodes in the distributed consensus network verify the signature validity and parameter digest trustworthiness using the corresponding public key, requiring a 100% pass rate for each verification. Reports that fail verification are discarded and a security alarm is triggered. After successful verification, other teaching nodes extract the parameter digest and transition thresholds to synchronize and update their local policy models. They can choose to request a complete incremental package of model parameters from the node via a secure differential synchronization channel, or perform incremental reconstruction based on the parameter digest and their local historical parameter baseline. Example

[0082] This embodiment was implemented in the campus culture construction scenario of the Department of Mechanical Engineering at a university in Gansu Province, involving two professional departments: CNC Technology and Industrial Robotics Technology, with a total of 6 teaching nodes (4 on-campus class nodes, 1 school-enterprise joint training base node, and 1 regional vocational education alliance shared node). Each node is deployed with a Trusted Execution Environment (TEE) based on Intel SGX, and cross-node privacy data collaborative computation is achieved through a distributed consensus network.

[0083] S1. Detrusted decoupling and encapsulation of campus culture-infused data manipulation intent declaration and multimodal privacy features. S1a. Data Acquisition and Mapping: Each teaching node collects raw, multi-source campus culture immersion data from September 2026 to March 2027 within its local TEE: Environmental and cultural immersion data: the time students spend in front of the digital craft display screen (average 127 seconds per visit), the cultural corridor visit trajectory (a total of 8,432 path points were recorded), and the training area environmental perception log. Activity cultural immersion data: Digital Craftsman Festival participation record (participation rate of CNC technology department 87.3%), Intelligent Design Week submission sequence (a total of 1,256 works), Mechanical Data Decoding Competition answer trajectory; AI-powered interactive immersion data: Dialogue records of the "Machine Companion" job cognition module (average 23.6 rounds per person per semester), frequency of ethical reminders triggered (cumulative 4,182 times).

[0084] Each node maps the three types of data to a unified education timeline through a 15-minute cultural immersion time window.

[0085] S1b. Generation and Usage Declaration of Program Data Operation Relationship Graph: The collaborative computing coordinator receives cross-node cultural immersion attribution computing task requests, performs static code analysis on the executable computing program, extracts the instruction sequences involving data access, generates a program data operation relationship graph, and distributes it to the 6 teaching nodes. Each node reviews the data usage according to its local preset strategy: the CNC technology department node prohibits the use of student training operation trajectories for strategy strength inference, and the school-enterprise joint node prohibits the use of enterprise process parameters for cross-node aggregation. After review, a first usage declaration certificate is generated, signed, and returned. The coordinator takes the intersection of the whitelist and the union of the blacklist to generate the collaborative computing task specification.

[0086] S1c. Feature Decoupling and Encapsulation: Each node inputs three types of data into the multimodal feature extraction model: the environmental perception embedding layer maps the cultural corridor visit trajectory into a 128-dimensional spatial semantic vector; the activity behavior encoding layer captures the temporal behavior patterns in the submission sequence of works from the Smart Design Week; and the interaction semantic compression layer transforms the dialogue records of the AI ​​companion into 64-dimensional semantic features. The outputs of the three layers are concatenated to form a 256-dimensional privacy-protected infiltrated feature vector. Simultaneously, the path coupling decoupling model calculates the interaction effect between each pair of the three paths: the interaction effect coefficient between environmental culture and activity culture is 0.34, the interaction effect coefficient between activity culture and AI interaction is 0.28, and the interaction effect coefficient between environmental culture and AI interaction is 0.19. After removing the main effects from the total effect, a coupling residual vector is generated. Verifiable credentials are generated through zero-knowledge proof. Differential privacy noise (ε=1.2) is applied to the feature vector and the coupling residual vector, and a data usage label is attached.

[0087] S2. Reliable Calculation of Cross-Node Cultural Immersion Net Attribution Contribution under Data Usage Constraints S2a. Cross-Domain Alignment and Construction of Policy Heterogeneity Subspace: Six teaching nodes publish privacy-preserving infiltration feature vectors with attached data usage labels, coupled residual vectors, and verifiable credentials to the distributed consensus network. After the collaborative nodes verify the verifiable credentials, the feature data of each node are mapped to a unified time coordinate system based on the temporal coupling relationship of the education timeline. In the environmental and cultural immersion dimension, the feature distribution difference between the school-enterprise joint node and the intra-school class node reaches 0.47 (exceeding the preset threshold of 0.35). In the AI ​​interaction dimension, the feature distribution difference between the regional alliance node and the intra-school node reaches 0.52. Based on this, a policy heterogeneity subspace is constructed.

[0088] S2b. Counterfactual Causal Inference and Attribution Calculation: Within the strategy heterogeneity subspace, collaborative nodes generate a virtual control strategy for each teaching node: reducing the intensity of environmental culture immersion by 30% while maintaining the intensity of activity culture and AI interaction. The attribution evaluation function is called to calculate the predicted values ​​of student digital literacy growth under the actual strategy and the virtual control strategy. The first predicted value for the CNC Technology Department node under the actual strategy is 0.78, and the second predicted value under the virtual control strategy is 0.61, with a node attribution contribution of 0.17. After weighted aggregation (weights normalized according to local sample size and strategy execution duration), the net cross-node attribution contribution of the environmental culture immersion path is 0.154 (15.4%), the activity culture immersion path is 0.231 (23.1%), and the AI ​​interaction immersion path is 0.312 (31.2%). Based on the dispersion of node attribution contributions on different paths (standard deviation 0.089), a strategy heterogeneity calibration factor of 1.18 is generated.

[0089] S2c. Purpose Execution Audit and Blocking: During the attribution computation process, the distributed consensus network monitors feature dimension access requests from collaborating nodes in real time. When a regional alliance node initiates an access request to a feature dimension corresponding to the enterprise's process parameters, and the opcode type of this request is not in the unified opcode type whitelist, the distributed consensus network immediately blocks the request, generates a purpose execution audit record, and synchronizes it to all teaching nodes. A total of 23 unauthorized access requests were blocked in this collaborative computation (a total of 6 nodes, averaging approximately 3.8 times per node). Based on the average number of blocking attempts, the collaborative computation credibility score is 0.967.

[0090] S3. Policy Credibility Oriented Evolution Under Attribution Audit Collaborative Feedback and Critical Leap Perception S3a. Critical Leap Point Detection: Each teaching node invokes the campus culture immersion strategy model in its local TEE. Based on the collaborative computing credibility score (0.967), the cross-node net attribution contribution is weighted with confidence, yielding the weighted net attribution contribution: Environmental Culture Path 0.149, Activity Culture Path 0.223, AI Interaction Path 0.302. Using the education timeline as a benchmark, the verified privacy protection immersion feature vector is segmented and extracted using a 7-day sliding time window, and the temporal change rate is calculated. The temporal change rate is then element-wise weighted and corrected based on the nonlinear interaction effect coefficient in the coupled residual vector. When the corrected temporal change rate exceeds the locally dynamically adjusted leap threshold (calibrated to 0.15 using historical data) for three consecutive time windows, the starting time of the first window meeting the condition (November 14, 2026) is determined as the critical leap point for students' digital literacy development.

[0091] S3b. Strategy Bias Calculation and Directed Evolution: The first strategy prediction output generated by the strategy model within the nearest window before the critical transition point is 0.62, the second strategy prediction output within the nearest window after the critical transition point is 0.84, and the change in student digital literacy growth prediction is 0.22. The difference between this and the weighted AI interaction path cross-node net attribution contribution of 0.302 yields a strategy bias of 0.082. Multiplying the strategy bias by the reciprocal of the strategy heterogeneity calibration factor (1 / 1.18 = 0.847) and the collaborative computation credibility score of 0.967 yields a calibrated strategy bias of 0.067. Minimizing this calibrated strategy bias is the optimization objective. The gradient direction of the loss function with respect to the model parameters is calculated, generating a directed optimization gradient. After 18 gradient descent iterations, the strategy model parameters converge. The updated model increases the resource allocation weight of the AI ​​interaction path from 0.28 to 0.35 and optimizes the recommended dwell time of the digital craft display screen in the environmental culture path from 127 seconds to 156 seconds. The parameter summary and transition threshold of 0.15 are sent back to the distributed consensus network as additional payloads via a remote proof mechanism for verification by other nodes.

[0092] Comparison of effects data To verify the technical effect of the present invention, this embodiment was compared with three existing technical solutions under the same dataset and computing environment. The comparison results are shown in Table 1 below. Figure 2 As shown: Table 1: Comparison of Results Comparative analysis: Privacy and security: This invention achieves a 36.5 percentage point improvement in gradient inversion attack resistance rate and a 53.6 percentage point improvement in internal unauthorized access blocking rate compared to general federated learning through dual isolation of TEE and differential privacy, combined with verifiable credentials of zero-knowledge proof.

[0093] Attribution accuracy: The path coupling decoupling model of this invention improves the separation accuracy of multi-path coupling effects to 94.5%, which is about 3 times higher than the existing technology; the strategy heterogeneity calibration factor compresses the cross-node attribution bias to 3.8%, which is 27.4 percentage points lower than the traditional TEE scheme.

[0094] Collaborative efficiency: This invention adopts a metadata constraint mechanism at the feature vector level, and the overhead of cross-node single collaborative communication is only 7.9% of that of secure multi-party computation, while the attribution computation latency is reduced by 80.9% compared to secure multi-party computation.

[0095] Educational effect: The critical transition detection mechanism of this invention accurately identifies the critical node of students' digital literacy from quantitative to qualitative change, with a detection accuracy of 91.3%, and the number of strategy evolution convergence iterations is reduced by 60% compared with general federated learning.

[0096] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

[0097] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this specification. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for collaborative computation of privacy-preserving data in a trusted execution environment of a computer, characterized in that, include: Three types of immersion data are collected and mapped to a unified education timeline to generate a program data operation relationship diagram. The program data operation relationship diagram is reviewed, a first purpose declaration certificate is generated and signed, the intersection of the whitelist and the union of the blacklist are taken, and a collaborative computing task specification is generated. Extract privacy-preserving infiltration feature vectors and coupled residual vectors, generate verifiable credentials based on zero-knowledge proofs, and then perform differential privacy desensitization and add data usage labels; Collaborative nodes combine privacy-preserving infiltration feature vectors to construct a policy heterogeneity subspace; Counterfactual causal inference calculates cross-node net attribution contribution and strategy heterogeneity calibration factor; distributed consensus network blocks unauthorized access and generates usage execution audit records. Each teaching node detects the critical transition point for students' digital literacy development and calculates the strategy deviation between the predicted output and the net attribution contribution across nodes. An optimized gradient update model is generated by combining a policy heterogeneity calibration factor, and the parameter summary and transition threshold are backpropagated to the distributed consensus network.

2. The method for collaborative computation of privacy-preserving data in a trusted execution environment according to claim 1, characterized in that, The three types of immersion data collected are mapped to a unified educational timeline to generate a program data operation relationship diagram, specifically including: Each teaching node collects original multi-source campus culture immersion data in a trusted execution environment, and maps three types of data—environmental culture immersion data, activity culture immersion data, and AI interaction immersion data—to a unified education timeline through a culture immersion time window. The collaborative computing coordinator receives the original collaborative computing request submitted by the computing task initiator and parses out the computing task identifier, the list of participants, and the executable computing program from it. The collaborative computing coordinator performs static code analysis on the executable computing program, traverses the entire instruction sequence, identifies each instruction related to data access, and extracts the target data field identifier involved in the instruction and the corresponding opcode type from each identified instruction. The collaborative computing coordinator, based on the control flow relationship of the executable computing program, associates and arranges the extracted target data field identifiers and opcode types in the execution order, generates a program data operation relationship diagram, and sends it to each teaching node.

3. The method for collaborative computation of privacy-preserving data in a trusted execution environment according to claim 2, characterized in that, The process of reviewing the program data operation relationship diagram, generating a first-purpose declaration credential, and signing it includes: Each teaching node, based on its preset data usage strategy corresponding to its locally stored privacy data, performs a step-by-step comparison of the opcode types contained in each record of the program data operation relationship diagram involving the privacy data of that teaching node and their execution order relationship in the control flow. Based on the comparison results, each teaching node adds the allowed operation code types to the first operation code type whitelist and the prohibited operation code types to the second operation code type blacklist. It then marks the corresponding data field identifier and program location range for each operation code type in the whitelist and generates a first purpose declaration certificate. Each teaching node uses its own private key to digitally sign the first purpose declaration credential and returns the signed first purpose declaration credential to the collaborative computing coordinator.

4. The method for collaborative computation of privacy-preserving data in a trusted execution environment according to claim 3, characterized in that, The process of taking the intersection of the whitelist and the union of the blacklist to generate collaborative computing task specifications specifically includes: The collaborative computing coordinator verifies the digital signature of each first-purpose declaration credential one by one to confirm that each first-purpose declaration credential has not been tampered with during transmission and originates from the corresponding teaching node. The collaborative computing coordinator extracts the first opcode type whitelist and the second opcode type blacklist of each teaching node from the verified first purpose declaration credential. It then takes the intersection of the whitelists and the union of the blacklists to obtain the unified opcode type whitelist and the unified opcode type blacklist. The collaborative computing coordinator extracts the program position ranges marked by each teaching node for each opcode type from all verified first-purpose declaration credentials, performs intersection and union of all program position ranges, and obtains the unified program position range constraints corresponding to each opcode type. The collaborative computing coordinator encapsulates the unified opcode type whitelist, unified opcode type blacklist, unified program position range constraints corresponding to each opcode type, and executable computing program to generate collaborative computing task specifications, and sends them to each teaching node.

5. The method for collaborative computation of privacy-preserving data in a trusted execution environment according to claim 4, characterized in that, The extraction of privacy-preserving infiltration feature vectors and coupled residual vectors, the generation of verifiable credentials based on zero-knowledge proofs, and the subsequent differential privacy desensitization and addition of data usage labels specifically include: Each teaching node inputs environmental culture immersion data, activity culture immersion data, and AI interaction immersion data into the corresponding modal channel of the multimodal feature extraction model. After extracting the deep semantic features of each modality, feature fusion is performed to generate a privacy-preserving immersion feature vector. Each teaching node inputs three types of data into the path coupling and decoupling model, calculates the interaction effect between each pair of the three immersion paths, and separates the main effect from the total effect to extract the path coupling residual vector of the campus culture immersion strategy of each teaching node. Each teaching node generates a local proof within the trusted execution environment based on the generation process of privacy-preserving infiltrated feature vectors and coupled residual vectors, and uses a zero-knowledge proof algorithm to transform the local proof into a verifiable credential. Each teaching node applies differential privacy noise perturbation to the privacy-protected infiltration feature vector and the coupled residual vector, and then associates and binds the perturbed vector with the data usage label to complete the desensitization and encapsulation. The data usage label records the mapping association between each dimension and the target data field identifier, as well as the corresponding unified opcode type whitelist, unified opcode type blacklist, and program position range constraint.

6. The method for collaborative computation of privacy-preserving data in a trusted execution environment according to claim 1, characterized in that, The collaborative nodes combine privacy-preserving infiltration feature vectors to construct a policy heterogeneity subspace, specifically including: Each teaching node publishes the privacy-preserving infiltrated feature vector, coupled residual vector, and verifiable credentials to the distributed consensus network through a trusted data delivery protocol; The collaborative node performs zero-knowledge proof verification on the verifiable credential to confirm that the decoupled encapsulation and purpose declaration process corresponding to the verifiable credential is completed in a trusted execution environment and the result has not been tampered with. After verification, the collaborative node extracts the time dimension information from the privacy protection infiltration feature vector of each teaching node, and combines it with the coupling residual vector to map the feature data to a unified time coordinate system based on the temporal coupling relationship of the education time axis, thus completing cross-domain alignment. In the feature space where cross-domain alignment is completed, the collaborative nodes calculate the feature distribution differences of each teaching node in the dimensions of environmental culture immersion, activity culture immersion, and AI agent interaction immersion. The feature offset vectors with differences exceeding a preset threshold are aggregated to construct a strategy heterogeneity subspace.

7. The method for collaborative computation of privacy-preserving data in a trusted execution environment according to claim 6, characterized in that, The counterfactual causal inference calculation of cross-node net attribution contribution and strategy heterogeneity calibration factor specifically includes: In the heterogeneous subspace of the strategy, the collaborative nodes generate a virtual comparison strategy based on the actual combination of strategy parameters adopted by each teaching node and according to the preset comparison strategy construction rules. The virtual comparison strategy only changes the strategy strength or strategy timing of the target infiltration path. Based on the actual strategies adopted by each teaching node and the virtual control strategy, the collaborative node calls the attribution evaluation function to calculate the first predicted value of students' digital literacy growth under the actual strategy and the second predicted value under the virtual control strategy, and calculates the difference between the two as the node attribution contribution on the corresponding target immersion path. The collaborative nodes perform weighted aggregation of the node attribution contributions of all teaching nodes on the same target immersion path to obtain the cross-node net attribution contribution of the target immersion path; Based on the dispersion of the node attribution contribution of each teaching node on different target immersion paths, the collaborative nodes calculate the influence weight of the differences in campus culture immersion strategies among teaching nodes on the attribution results, and generate a strategy heterogeneity calibration factor.

8. The method for collaborative computation of privacy-preserving data in a trusted execution environment according to claim 7, characterized in that, The distributed consensus network blocks unauthorized access and generates an audit log for the purpose of execution, specifically including: During the attribution calculation process, the distributed consensus network monitors every feature dimension access request initiated by the cooperating nodes in real time and extracts the target dimension identifier and request opcode type from the feature dimension access request. The distributed consensus network uses the target dimension identifier as an index to query the data usage tag attached to the dimension, and parses the corresponding Unicode type whitelist, Unicode type blacklist and program location range constraint from the data usage tag. The distributed consensus network determines whether the request opcode type belongs to the unified opcode type whitelist and not to the unified opcode type blacklist. If not, it determines that the request exceeds the scope declared by the collaborative computing task specification and immediately blocks the access request for that feature dimension. After completing the blocking, the distributed consensus network obtains the teaching node identifier, target feature dimension identifier, request operation code type and blocking timestamp involved in the blocked request, generates a purpose execution audit record, saves it to the audit log storage area and synchronizes it to all teaching nodes.

9. The method for collaborative computation of privacy-preserving data in a trusted execution environment according to claim 1, characterized in that, Each teaching node detects the critical transition point in students' digital literacy development, and calculates the strategy deviation between the strategy prediction output and the cross-node net attribution contribution, specifically including: Each teaching node invokes the local campus culture immersion strategy model, executes audit records to count the number of blocked requests and calculates the collaborative computing credibility score according to the purpose; based on the score, the confidence of the cross-node net attribution contribution is weighted, the availability of the privacy protection immersion feature vector is verified, and feature dimensions that cannot be published due to local policy restrictions are masked to generate a verified vector; based on the verified vector and the coupled residual vector, the critical transition point for the growth of students' digital literacy is detected. Each teaching node uses the education timeline as a benchmark, extracts the verified privacy protection infiltration feature vector in segments according to the sliding time window, calculates the temporal change rate within each window, calculates the nonlinear interaction effect intensity coefficient based on the L2 norm of the coupled residual vector, performs scalar weighted correction on the temporal change rate, compares the corrected temporal change rate with the locally dynamically adjusted transition threshold, and determines the critical transition point when the corrected temporal change rate exceeds the transition threshold in multiple consecutive windows. Each teaching node obtains the first and second strategy prediction outputs generated by the campus culture immersion strategy model based on local feature input within the most recent time window before and after the critical transition point, calculates the difference between the two, and obtains the predicted change in students' digital literacy growth. Each teaching node selects the weighted net attribution contribution of the target immersion path corresponding to the current strategy optimization goal. The difference between the predicted change in students' digital literacy growth and this contribution is calculated to obtain the strategy deviation between the strategy prediction output of the campus culture immersion strategy model before and after the critical transition point and the net attribution contribution of the cross-node.

10. A method for collaborative computation of privacy-preserving data in a trusted execution environment according to claim 9, characterized in that, The process of generating an optimized gradient update model by combining a strategy heterogeneity calibration factor and then feeding back the parameter summary and transition threshold to the distributed consensus network specifically includes: Each teaching node multiplies the policy bias by the reciprocal of the policy heterogeneity calibration factor, and then multiplies it by the collaborative computation credibility score to obtain the calibrated policy bias after eliminating cross-node policy difference interference and correcting the collaborative computation credibility. Each teaching node takes minimizing the deviation of the calibration strategy as the optimization objective, calculates the gradient direction of the loss function with respect to the model parameters under the current parameters of the campus culture immersion strategy model, and generates a directional optimization gradient; Each teaching node updates the current parameters of the local campus culture immersion strategy model by gradient descent according to the directional optimization gradient, obtains the updated strategy model parameters, and extracts and splices them into a parameter summary by dimensionality reduction. The transition thresholds corresponding to the critical transition points are merged with the parameter summary as independent parameters to construct the data to be transmitted back. Each teaching node invokes the remote proof module of its local trusted execution environment, uses the data to be transmitted back as the additional payload of the remote proof report, generates a hardware-signed remote proof report, and transmits it back to the distributed consensus network for other teaching nodes to verify the credibility of the updated policy model parameters.

Images