A data protection method for an online mall system based on a blockchain

Abstract

The application discloses a kind of data protection methods of online mall system based on blockchain, comprising the following steps: online mall transaction action, business stage, sensitive attribute commitment, contract verification relationship and external access request are written into action query end;Heterogeneous message is decomposed into business necessary message and privacy inversion message;Locate attention head and relationship edge affected by sensitive disturbance sample;Perform null space projection;Generate disturbance invariance gate result;Accumulate inversion budget consumption value, control business action result granularity;Write into blockchain, execute smart contract release check.The application relates to the technical field of online mall data security, realizes the trusted protection of online mall data, reduces the privacy inversion risk, and improves the reliability of on-chain audit.

Description

Technical Field

[0001] This invention relates to the field of online shopping mall data security technology, and in particular to a data protection method for an online shopping mall system based on blockchain. Background Technology

[0002] Online e-commerce systems generate data such as transaction actions, account status, address associations, risk tags, and contract verification records during transaction processing, payment verification, refund approval, after-sales processing, and audit verification. Existing data protection solutions mostly employ field anonymization, access control, encrypted storage, blockchain notarization, and smart contract authorization to restrict access to sensitive fields, record the flow of transaction data, and prevent unauthorized access to business data.

[0003] Existing solutions primarily focus on protecting the data fields themselves, lacking fine-grained constraints on the propagation paths during model inference. When using graph models or attention models to handle transaction actions and contract verification relationships in online e-commerce systems, the attention head aggregates information from business stages, sensitive attribute commitments, and transaction relationships. If different relationship propagation components directly enter the business action result output layer, even if the business action result does not directly output sensitive fields, sensitive attribute clues may still be exposed due to category changes, excessively fine-grained output, or the direction of relationship edge propagation, creating a privacy reversal risk.

[0004] Existing blockchain-based approval verification is mostly used to record authorization results or verify transaction status, making it difficult to verify whether the model has completed privacy-preserving message separation, perturbation response detection, null-space projection, and output perturbation invariance control. Once business action results are directly output through the ordinary authorization process, there is a lack of unified constraints on class flipping, inversion budget consumption, and changes in propagation weights under sensitive perturbations. This makes it difficult for online e-commerce systems to balance maintaining business processing efficiency and suppressing privacy-preserving inversion.

[0005] Therefore, how to provide a data protection method for a blockchain-based online shopping mall system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a data protection method for a blockchain-based online marketplace system. This invention utilizes heterogeneous graph Transformer, relation-aware attention decomposition, null projection, and smart contract shadow replay technology to perform privacy inversion suppression and on-chain release verification on the output process of online marketplace business actions. It has the advantages of controllable business output, strong privacy protection, and reliable audit traceability.

[0007] A data protection method for a blockchain-based online marketplace system according to an embodiment of the present invention includes the following steps: S1. Embed online mall transaction actions, business stages, sensitive attribute commitments and contract verification relationships into the action query end of the heterogeneous graph Transformer, and write external access requests into the action query end. S2. Using the relation-aware attention decomposition operator, the heterogeneous messages received by the action query terminal are decomposed into business-necessary messages and privacy-reversal messages. S3. Apply sensitive perturbation sample response detection to privacy inversion messages to locate attention heads and relation edges that cause changes in business action results with sensitive perturbation samples; S4. Perform null-spatial projection on the attention heads and relation edges obtained from the positioning to reduce the propagation weight of privacy inversion messages and retain the propagation weight of business-necessary messages. S5. Inject the necessary business messages after null space projection back into the action query terminal, perform perturbation invariance gating on the business action results, and obtain perturbation invariance gating results, retaining only the business action results that do not undergo class flipping with sensitive perturbation samples. S6. Accumulate the inversion budget consumption value according to the exclusion ratio of sensitive disturbance samples based on the retained business action results. When the inversion budget consumption value reaches the threshold, reduce the granularity of the business action results. If the threshold is still reached after reducing the granularity, terminate the output of the business action results. S7. Write the null space projection result, perturbation invariance gating result, and inversion budget consumption value into the blockchain, execute the release verification through smart contract, output the business action result after the verification passes, and write the abnormal audit record when the verification fails.

[0008] Optionally, S2 specifically includes: S21. Disconnect the multi-head attention aggregation link between the action query end and the business action result output layer in the heterogeneous graph Transformer, and cut off the connection that heterogeneous messages directly enter the business action result output layer after being weighted by attention heads. S22. In each attention head, a relation-aware attention decomposition operator is connected to perform relation projection, business stage masking, and action consistency judgment on each heterogeneous message in sequence. S23. When the heterogeneous message after relation projection passes through the business stage mask and is judged by action consistency, the output of the corresponding attention head is written into the necessary business message. S24. When the heterogeneous message after relation projection fails any judgment in the business stage mask or action consistency judgment, write the output of the corresponding attention head into the privacy inversion message. S25. Retain the propagation connection to the business action result output layer for business-necessary messages, and close the propagation connection to the business action result output layer for privacy inversion messages. S26. Use business-necessary messages and privacy-inverting messages as the decomposition output of the relation-aware attention decomposition operator.

[0009] Optionally, the consistency determination of service stage mask and action in S22 specifically includes: S221. Input the relation projection result into the business stage mask, and control the attention head propagation connection according to the current business stage. S222. When the relation projection result falls within the processing scope of the current business stage, open the propagation connection from the corresponding attention head to the action consistency judgment. S223. When the relation projection result is outside the processing scope of the current business stage, the corresponding attention head will be output bypassed to the privacy inversion message. S224. After the attention head output is enabled by the business stage mask, it is connected to the local attention window of the action query terminal, and the semantic offset comparison before and after the connection is performed. S225. When the attention head output does not change the action consistency judgment result and does not trigger the sensitive attribute commitment to participate in the local attention window aggregation, the corresponding attention head output will be written into the business necessary message. S226. When the attention head output changes the action consistency judgment result or triggers a sensitive attribute commitment to participate in local attention window aggregation, the corresponding attention head output is written into the privacy inversion message.

[0010] Optionally, the sensitive perturbation sample response detection applied to the privacy inversion message in step S3 specifically includes: S31. Freeze the propagation weight of business-necessary messages to the business action result output layer, and close the connection of business-necessary messages to the sensitive disturbance sample response detection loop. S32. Integrate the privacy inversion message into the sensitive disturbance sample response detection loop, and apply paired sensitive disturbance samples with opposite directions and consistent amplitudes within the same attention head; S33. Keep the action query end, business stage and relationship type unchanged, lock the attention head transmission order and relationship edge propagation path of the pair of sensitive disturbance samples, and transmit the pair of sensitive disturbance samples to the business action result output layer respectively. S34. Execute response feedback from the business action result output layer to each attention head, and lock the perturbation sensitive attention head according to the output difference caused by the paired sensitive perturbation samples. S35. Within the perturbation-sensitive attention head, mask each relation edge one by one, compare the output difference before and after masking, and lock the masked relation edge as the perturbation-sensitive relation edge when the output difference decreases to the elimination threshold. S36. Add and compare the two sets of opposite output differences corresponding to the perturbation-sensitive attention head and the perturbation-sensitive relation edge. When the addition result drops to the cancellation threshold, the perturbation-sensitive attention head and the perturbation-sensitive relation edge are identified as the attention head and relation edge that cause the business action result to change with the sensitive perturbation sample.

[0011] Optionally, in step S4, null-space projection is performed on the attention head and relation edges obtained from the localization, specifically including: S41. Lock the perturbation-sensitive relation edge in the perturbation-sensitive attention head, and import the privacy inversion message corresponding to the perturbation-sensitive relation edge into the null space projection link; S42. Within the sensitive attention head of the disturbance, disconnect the shared aggregation entry of privacy inversion messages and business-necessary messages, and import the privacy inversion messages separately into the inversion direction stripping entry; S43. Using the propagation direction of the perturbation sensitive relation edge as the null space constraint direction, perform orthogonal suppression on the privacy inversion message in the inversion direction stripping entry; S44. Feed the privacy inversion message after orthogonal suppression back to the output side of the perturbation sensitive attention head, and retain the propagation weight of business-necessary messages in the perturbation sensitive attention head; S45. Perform post-projection weight reset on the perturbation-sensitive relation edge and close the attention weight transfer path from the perturbation-sensitive relation edge to the sensitive attribute commitment direction; S46. Before the perturbation-sensitive attention head completes its output, the propagation weights of the business-necessary messages and the privacy-inverting messages after orthogonal suppression are weighted and merged. The original propagation path of the business-necessary messages is retained, and the privacy-inverting messages after orthogonal suppression are limited to null-space projection results to participate in the weight merging.

[0012] Optionally, obtaining the perturbation invariance gating result in S5 specifically includes: S51. Inject the necessary business messages after zero-space projection back into the action query terminal, and lock the action query terminal, business stage, relationship type and attention head transmission order. S52. Within the disturbance invariance gating, sensitive disturbance samples are closed and opened respectively, and the gating baseline category and business action result category are obtained according to the locked attention head transmission order. S53. When the business action result category is consistent with the gating benchmark category, the output connection of the business action result output layer is retained. S54. When the business action result category is flipped relative to the gating baseline category, the output connection of the business action result output layer is cut off, and the corresponding business action result output is blocked. S55. Write the corresponding states in the output connection retention state and output connection disconnection state of the business action result output layer into the disturbance invariant gating result.

[0013] Optionally, the cumulative inversion budget consumption value in S6 specifically includes: S61. Write the retained business action result categories in reverse to the sensitive perturbation samples corresponding to the perturbation sensitive attention head and the perturbation sensitive relation edge, and mark the sensitive perturbation samples excluded by the business action result categories. S62. Accumulate and invert the budget consumption value by taking the marked sensitive disturbance sample along the budget back pressure gate and the corresponding disturbance sensitive attention head and disturbance sensitive relation edge. S63. When the inversion budget consumption value does not reach the threshold, maintain the current output connection of the business action result output layer; S64. When the inversion budget consumption value reaches the threshold, the business action result category is converted into the higher-level granular business action result category. S65. After the business action result category is converted, if the budget back pressure gate accumulates and inverts the budget consumption value again and still reaches the threshold, the output fuse gate cuts off all output connections of the business action result output layer and terminates the business action result output.

[0014] Optionally, the release verification performed in S7 via a smart contract specifically includes: S71. Cut off the direct connection between the business action result output layer and the external output end within the chain certificate release layer, and connect the business action result output layer to the smart contract shadow replay channel. S72. Write the null space projection results, perturbation invariance gating results, and inversion budget consumption value into the blockchain; S73, The smart contract shadow replay channel performs shadow replay on the null space projection result, the disturbance invariance gating result and the inversion budget consumption value to generate the first release flag, the second release flag and the third release flag; S74. After the first release flag, the second release flag, and the third release flag are all written, the release latch connects the business action result output layer with the external output terminal. S75. If any release flag is not written, the release latch keeps the business action result output layer disconnected from the external output terminal and writes the shadow replay result corresponding to the missing release flag into the abnormal audit record.

[0015] Optionally, the shadow replay performed in S73 specifically includes: S731, the smart contract shadow replay channel includes the projection weight reset verification path, the category flip verification path, and the budget circuit breaker verification path; S732, read the null space projection result of the post-projection weight reset verification path, replay the post-projection weight reset action, and write the first release flag when the attention weight transfer path from the perturbation sensitive relation edge to the sensitive attribute commitment direction remains closed. S733, the category flip verification path reads the disturbance invariance gating result, replays the retained output connection action, and writes the second release flag when the business action result category has not been flipped relative to the gating base category. S734, the budget circuit breaker verification path reads the inverted budget consumption value, replays the budget back pressure gate and output circuit breaker gate actions, and writes the third release flag when the output circuit breaker gate is not triggered.

[0016] Optionally, the release latch in S74 includes a first release tag latch bit, a second release tag latch bit, a third release tag latch bit, and an output connection control bit. When all three release tag latch bits are in the write state, the output connection control bit opens the output connection between the service action result output layer and the external output terminal. When any release tag latch bit is not written, the output connection control bit closes the output connection between the service action result output layer and the external output terminal. After the service action result is output, the three release tag latch bits are cleared.

[0017] The beneficial effects of this invention are: (1) This invention sets a relation-aware attention decomposition operator on the attention head output side of the heterogeneous graph Transformer to decompose heterogeneous messages into business-necessary messages and privacy-reversal messages, cutting off the path for undecomposed heterogeneous messages to directly enter the business action result output layer. Compared with field desensitization, permission authorization, and on-chain evidence storage, this invention can intercept the propagation components of sensitive attribute commitment direction during the model propagation stage, avoiding the business action result from still carrying privacy-reversal clues when no sensitive fields are output.

[0018] (2) This invention applies paired sensitive perturbation samples to the privacy inversion message, combines response backflow and relation edge masking, locks the perturbation sensitive attention head and perturbation sensitive relation edge, and then uses null projection to suppress the propagation weight of the privacy inversion message in the sensitive attribute commitment direction. This processing is not a post-processing of the output result, but a targeted weakening of the privacy inversion propagation path at the attention head and relation edge level, which can reduce the risk of class flipping while retaining the propagation weight of business-necessary messages.

[0019] (3) This invention writes the null projection results, perturbation invariance gating results, and inversion budget consumption values ​​into the blockchain. The smart contract shadow replay channel verifies the weight reset action after projection, the output connection retention action, the budget backpressure gate, and the output circuit breaker action, respectively. Then, the release latch controls the connection status between the business action result output layer and the external output terminal. This process extends the on-chain verification object from ordinary authorization records to privacy control actions within the model, preventing the direct output of business action results that have not completed privacy inversion suppression, and improving the reliability of release verification and the ability to trace anomalies. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a data protection method for a blockchain-based online marketplace system proposed in this invention. Figure 2 This is a schematic diagram of the relation-aware attention decomposition and sensitive perturbation response detection proposed in this invention; Figure 3 This is a schematic diagram of the null space projection, perturbation invariance gating, and chain release proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 A data protection method for a blockchain-based online marketplace system includes the following steps: S1. Embed online mall transaction actions, business stages, sensitive attribute commitments and contract verification relationships into the action query end of the heterogeneous graph Transformer, and write external access requests into the action query end. In this invention, the action query terminal of the heterogeneous graph Transformer is the model entry point for receiving external access requests within the heterogeneous graph Transformer. It consists of action class input slots, stage class input slots, commitment class input slots, and verification class input slots. During implementation, online mall transaction actions are written into action class input slots, business stages are written into stage class input slots, sensitive attribute commitments are written into commitment class input slots, and contract verification relationships are written into verification class input slots. The action query terminal is used to transform external access requests into business action queries, enabling subsequent relationship-aware attention decomposition operators to propagate and distribute around the business action query execution. In this invention, online mall transaction actions are obtained from the request purpose and business operation type in the external access request; during implementation, the external access request is parsed, and the business operation types in order placement, payment, delivery, refund, after-sales verification, and audit verification are mapped to online mall transaction actions; online mall transaction actions are used to determine the business action result category that the business action result output layer needs to judge. In this invention, the business stage is obtained from the processing process of the online mall transaction action; during implementation, the processing status of transaction creation, payment confirmation, logistics flow, refund processing, after-sales verification, and audit verification is read and written to the action query terminal; the business stage is used to limit the propagation range of the relationship projection result and prevent the attention head output that is out of the current business stage from entering the business action result output layer; In this invention, the sensitive attribute commitment is obtained by committing the sensitive attributes; in practice, the user identity attribute, transaction preference attribute, account risk attribute, and address association attribute are committed and encoded to form a verification object that does not expose the plaintext of the sensitive attributes; the sensitive attribute commitment is used to participate in local attention window aggregation checks, sensitive disturbance sample generation, and smart contract release verification. In this invention, the contract verification relationship is obtained from the correspondence between the release conditions in the smart contract, on-chain records, and model outputs; during implementation, the contract verification relationship is established according to the verification requirements of the null space projection results, the perturbation invariance gating results, and the inversion budget consumption value; the contract verification relationship is used to support the on-chain verification of the internal control actions of the model by the smart contract shadow replay channel; In this invention, writing external access requests to the action query terminal is an operation that transforms access requests into business action queries within the model. In practice, the request subject, request purpose, and business action type are extracted from the external access request, and the business action type is written to the action query terminal without directly triggering the reading of sensitive fields. After writing is completed, the action query terminal sends the business action query into the attention propagation link of the heterogeneous graph Transformer. S2. Using the relation-aware attention decomposition operator, the heterogeneous messages received by the action query terminal are decomposed into business-necessary messages and privacy-reversal messages. In this invention, heterogeneous messages are obtained by attention propagation of different relation types within the heterogeneous graph Transformer; in practice, transaction action relations, business stage relations, sensitive attribute commitment relations, and contract verification relations are propagated through their respective attention heads to form heterogeneous messages with relation sources; the heterogeneous messages are used to enter the relation-aware attention decomposition operator as decomposition objects for business-necessary messages and privacy-reversal messages; In this invention, the relation-aware attention decomposition operator is a propagation splitting structure set between the output side of the attention head and the input side of the business action result output layer. In implementation, the relation-aware attention decomposition operator is connected in each attention head to sequentially perform relation projection, business stage masking, and action consistency judgment on heterogeneous messages. The relation-aware attention decomposition operator is used to write the propagation components that can support the output of business action results into the necessary business messages and write the propagation components that may cause sensitive attribute inversion into the privacy inversion messages. In this invention, the necessary business messages are obtained by the output of the attention head through the business stage mask and action consistency judgment. In practice, when the relation projection result falls within the processing range of the current business stage and the attention head output does not change the action consistency judgment result, the corresponding attention head output is written into the necessary business messages. The necessary business messages are used to enter the business action result output layer to support the generation of business action result categories. In this invention, the privacy inversion message is obtained from the output of the attention head that fails to pass the business stage mask, fails the action consistency judgment, or triggers the sensitive attribute commitment to participate in the local attention window aggregation; in implementation, the above attention head output is bypassed from the input side of the business action result output layer and written into the privacy inversion message; the privacy inversion message is used to enter the sensitive disturbance sample response detection loop and locate the privacy inversion propagation path; S3. Apply sensitive perturbation sample response detection to privacy inversion messages to locate attention heads and relation edges that cause changes in business action results with sensitive perturbation samples; In this invention, sensitive disturbance sample response detection is an operation to detect whether privacy inversion messages affect the results of business actions. In practice, the connection of necessary business messages to the sensitive disturbance sample response detection loop is closed, the privacy inversion messages are connected to the detection loop, and pairs of sensitive disturbance samples with opposite directions and consistent amplitudes are applied at the latent vector position corresponding to the sensitive attribute commitment. Then, they are propagated to the business action result output layer along the same attention head and the same relation edge, and the changes in the business action results are compared. In this invention, the attention head that causes the business action result to change with the sensitive disturbance sample is obtained by the response feedback; in practice, the output difference caused by the sensitive disturbance sample is fed back from the business action result output layer to each attention head, and the attention head whose output difference response reaches the set condition is locked as the disturbance sensitive attention head; the disturbance sensitive attention head is used as the position of action for subsequent relation edge screening and null space projection; In this invention, the relational edges that cause the business action results to change with the sensitive perturbation samples are obtained by masking each relational edge one by one; in practice, each relational edge is masked within the perturbation sensitive attention head, and the output difference before and after masking is compared; when the output difference drops to the elimination threshold, the masked relational edge is locked as the perturbation sensitive relational edge; the perturbation sensitive relational edge is used to determine the null space constraint direction and the weight reset object after projection; S4. Perform null-spatial projection on the attention heads and relation edges obtained from the positioning to reduce the propagation weight of privacy inversion messages and retain the propagation weight of business-necessary messages. In this invention, null-space projection is an operation that weakens the propagation component of the sensitive attribute commitment direction in the privacy inversion message. In implementation, the propagation direction of the perturbed sensitive relationship edge is used as the null-space constraint direction. The privacy inversion message corresponding to the perturbed sensitive relationship edge is imported into the null-space projection link. Orthogonal suppression is performed on the privacy inversion message, and the weight is reset after projection on the perturbed sensitive relationship edge. Null-space projection is used to reduce the propagation weight of the privacy inversion message while retaining the propagation weight of the necessary business message. S5. Inject the necessary business messages after null space projection back into the action query terminal, perform perturbation invariance gating on the business action results, and obtain perturbation invariance gating results, retaining only the business action results that do not undergo class flipping with sensitive perturbation samples. In this invention, injecting the necessary business messages after null-space projection back into the action query terminal is an operation that sends the business propagation components after projection back to the action query entry point. During implementation, the order of action query terminal, business stage, relationship type, and attention head transmission remains unchanged, and the necessary business messages after null-space projection are sent back to the action query terminal. The injection operation is used to provide comparison input under the same propagation conditions for disturbance invariance gating. In this invention, disturbance invariance gating is an output control operation used to determine whether the result of a business action is affected by a sensitive disturbance sample. During implementation, sensitive disturbance samples are turned off and on within the disturbance invariance gating, and the gating baseline category and the business action result category are obtained according to the locked attention head transmission order. The output connection is maintained when the business action result category matches the gating baseline category, and the output connection is severed when a category flip occurs. The disturbance invariance gating is used to block the output of business action results that change with the sensitive disturbance sample. In this invention, the perturbation invariance gating result is obtained after the perturbation invariance gating completes the category comparison; during implementation, the corresponding states in the output connection retention state and output connection disconnection state of the business action result output layer are written into the perturbation invariance gating result; the perturbation invariance gating result is used to write into the blockchain for category flip verification path review and retention of output connection action; S6. Accumulate the inversion budget consumption value according to the exclusion ratio of sensitive disturbance samples based on the retained business action results. When the inversion budget consumption value reaches the threshold, reduce the granularity of the business action results. If the threshold is still reached after reducing the granularity, terminate the output of the business action results. In this invention, the inversion budget consumption value is accumulated from the ability of the business action result category to exclude sensitive disturbance samples. During implementation, the retained business action result categories are written in reverse to the sensitive disturbance samples corresponding to the disturbance sensitive attention head and the disturbance sensitive relationship edge. Sensitive disturbance samples excluded by the business action result categories are marked, and the budget back pressure gate accumulates the inversion budget consumption value along the path corresponding to the marked sensitive disturbance samples. The inversion budget consumption value is used to determine whether the business action result has an excessively strong inversion capability. In this invention, reducing the granularity of business action results involves converting fine-grained business action result categories into higher-level granular business action result categories. In practice, when the inversion budget consumption value reaches a threshold, the output of fine-grained business action result categories is turned off, and the business action result categories are converted into higher-level granular business action result categories. Reducing the granularity of business action results is used to decrease the ability of the output results to exclude sensitive disturbance samples. In this invention, the termination of business action result output is an output control operation executed when the inversion budget consumption value still reaches the threshold after reducing the granularity; during implementation, the budget back pressure gate accumulates the inversion budget consumption value again; when the result of the re-accumulation still reaches the threshold, the output circuit breaker gate cuts off all output connections of the business action result output layer; the termination of business action result output is used to prevent the business action result from continuing to release privacy inversion clues; S7. Write the null space projection result, perturbation invariance gating result, and inversion budget consumption value into the blockchain, execute the release verification through smart contract, output the business action result after the verification passes, and write the abnormal audit record when the verification fails.

[0023] In this invention, the null projection result is obtained by the null projection link completing orthogonal suppression and post-projection weight reset; during implementation, the disturbance-sensitive attention head, disturbance-sensitive relation edge, null constraint direction, orthogonal suppression state and post-projection weight reset state are recorded to form the null projection result; the null projection result is used to write to the blockchain for the smart contract shadow replay channel to verify the post-projection weight reset action.

[0024] In this invention, the smart contract execution release verification is an on-chain review operation before the business action result is output. During implementation, the null space projection result, the perturbation invariance gating result, and the inversion budget consumption value are written into the blockchain. The smart contract shadow replay channel reads the three types of results and performs shadow replay to generate the first release flag, the second release flag, and the third release flag. The smart contract release verification is used to determine whether the business action result has the conditions for external output.

[0025] In this invention, the anomaly audit record is obtained when the smart contract release verification fails. In practice, when any of the first, second, and third release flags is not written, the release latch keeps the business action result output layer disconnected from the external output terminal. The smart contract shadow replay channel reads the shadow replay result corresponding to the missing release flag and writes the business stage, business action result category, missing release flag type, and failure reason into the anomaly audit record. The anomaly audit record is used to trace the reason for the on-chain release failure in the future.

[0026] In this embodiment, S2 specifically includes: S21. Disconnect the multi-head attention aggregation link between the action query end and the business action result output layer in the heterogeneous graph Transformer, and cut off the connection that heterogeneous messages directly enter the business action result output layer after being weighted by attention heads. In this invention, the business action result output layer is a model structure in the heterogeneous graph Transformer that outputs business action result categories, consisting of an action classification output unit and an output connection control unit. In practice, the business action result output layer only receives the necessary business messages output by the relation-aware attention decomposition operator, and does not receive undecomposed heterogeneous messages or propagation components that are judged as privacy inversion messages. The business action result output layer is used to generate business action result categories in order release, payment verification, refund review, after-sales verification, and audit verification based on the necessary business messages, and serves as the output object for disturbance invariance gating, inversion budget control, and smart contract release verification. In this invention, the multi-head attention aggregation link is the internal propagation path of the model between the action query end and the business action result output layer; this link is formed by the weighted output of several attention heads and is used to transmit the heterogeneous messages received by the action query end to the business action result output layer; this invention sets a splitting position in this link so that the output of the attention head first enters the relation-aware attention decomposition operator; In this invention, cutting off the connection between heterogeneous messages and the business action result output layer is to block the undecomposed attention head output from directly participating in the generation of business action results. In practice, the direct connection path from the attention head output side to the input side of the business action result output layer is closed, the attention head output is transferred to the relation-aware attention decomposition operator, and the decomposition result determines whether to enter the business action result output layer. S22. In each attention head, a relation-aware attention decomposition operator is connected to perform relation projection, business stage masking, and action consistency judgment on each heterogeneous message in sequence. In this invention, the operation of setting the propagation splitting structure on the output side of the attention head is to integrate the relation-aware attention decomposition operator into each attention head; in practice, after each attention head completes the weighted output, it sends the corresponding heterogeneous message into the decomposition entry, and the decomposition entry performs subsequent discrimination according to the relation source and the current business stage; In this invention, relation projection, business stage masking, and action consistency judgment are executed sequentially. During implementation, relation projection first maps heterogeneous messages to corresponding propagation paths according to the transaction action relationships, business stage relationships, sensitive attribute commitment relationships, or contract verification relationships carried by the heterogeneous messages. The business stage mask then performs on-screen, suppression, or bypass control on the corresponding propagation path based on the current business stage. Attention head outputs that fall within the processing scope of the current business stage enter the action consistency judgment, while attention head outputs that fall outside the processing scope of the current business stage are bypassed to the privacy inversion message. The action consistency judgment connects the attention head outputs through the business stage mask to the local attention window of the action query end, compares the semantic changes of business actions before and after the connection, and writes the attention head outputs that do not change the action consistency judgment result into the necessary business message. Attention head outputs that change the action consistency judgment result or trigger sensitive attribute commitments to participate in aggregation are written into the privacy inversion message. S23. When the heterogeneous message after relation projection passes through the business stage mask and is judged by action consistency, the output of the corresponding attention head is written into the necessary business message. S24. When the heterogeneous message after relation projection fails any judgment in the business stage mask or action consistency judgment, write the output of the corresponding attention head into the privacy inversion message. S25. Retain the propagation connection to the business action result output layer for business-necessary messages, and close the propagation connection to the business action result output layer for privacy inversion messages. In this invention, after the relation-aware attention decomposition operator completes the business stage mask and action consistency judgment, it rewrites the propagation state of the attention head output side according to the judgment result. In practice, the original propagation weight and original propagation path are retained for the attention head output written into the business necessary message, and the connection mark leading to the business action result output layer is deleted for the attention head output written into the privacy inversion message. The subsequent propagation position of the attention head output is switched to the sensitive disturbance sample response detection loop. This process is used to complete the propagation path rewriting at the attention head output side, rather than performing filtering after the business action result is output. S26. Use business-necessary messages and privacy-inverting messages as the decomposition output of the relation-aware attention decomposition operator.

[0027] In this embodiment, the consistency determination of service stage mask and action in S22 specifically includes: S221. Input the relation projection result into the business stage mask, and control the attention head propagation connection according to the current business stage. In this invention, the business stage mask is an internal control operation of the model that controls the attention head propagation connection based on the current business stage. In practice, after the relation projection result enters the business stage mask, the business stage mask reads the current business stage and matches the propagation connection corresponding to the relation projection result with the processing range of the current business stage. The matching result is used to determine whether the attention head propagation connection enters the open state, the suppressed state, or the bypass state. In this invention, the current business stage is obtained from the business stage written in the action query terminal; in practice, the business stage in the action query terminal is used as the judgment benchmark for the business stage mask. The business stage mask does not regenerate the business stage, but only reads the current business stage and controls the attention head propagation connection accordingly; the current business stage is used to limit whether the attention head output allows continued entry into the action consistency judgment. S222. When the relation projection result falls within the processing scope of the current business stage, open the propagation connection from the corresponding attention head to the action consistency judgment. In this invention, the processing scope of the current business stage is the range of relationships that are allowed to participate in the judgment of business action results under the current business stage. In practice, based on the business stage written in the action query terminal, the transaction action relationships, business stage relationships, sensitive attribute commitment relationships, and contract verification relationships that are from the same source as the business stage are read, and the relationship propagation range that does not match the current business stage is excluded to form the processing scope of the current business stage. The processing scope of the current business stage is used as the matching basis for the business stage mask to filter out relationship projection results that are irrelevant to the current business stage. S223. When the relation projection result is outside the processing scope of the current business stage, the corresponding attention head will be output bypassed to the privacy inversion message. In this invention, bypassing to the privacy inversion message is an operation that removes the attention head output that is out of the processing scope of the current business stage from the input direction of the business action result output layer. In practice, after the projection result of the business stage mask judgment relationship does not belong to the processing scope of the current business stage, the propagation connection of the attention head to the action consistency judgment is no longer opened, and the corresponding attention head output is written into the privacy inversion message. This operation is used to prevent the propagation components across business stages from participating in the generation of business action results. S224. After the attention head output is enabled by the business stage mask, it is connected to the local attention window of the action query terminal, and the semantic offset comparison before and after the connection is performed. In this invention, the local attention window is the calculation range within the action query terminal used to compare the semantic changes of business actions before and after the attention head output is connected. In practice, the action consistency discrimination reads the attention head output after it has been enabled by the business stage mask and connects the attention head output to the local attention window of the action query terminal. Before connection, the semantic state of the business actions in the action query terminal is recorded. After connection, the semantic state of the business actions is read again, and the difference between the states before and after connection is used to determine whether the attention head output changes the action consistency discrimination result. S225. When the attention head output does not change the action consistency judgment result and does not trigger the sensitive attribute commitment to participate in the local attention window aggregation, the corresponding attention head output will be written into the business necessary message. In this invention, not changing the action consistency judgment result means that after the attention head output is connected to the local attention window, the semantic state of the business action does not deviate from the online mall transaction action written in the action query terminal. In practice, the action consistency judgment compares the semantic state of the business action before and after the connection. If the attention head output does not change the action consistency judgment result and does not trigger the sensitive attribute commitment to participate in the local attention window aggregation, then the corresponding attention head output is written into the necessary business message. The above processing is used to retain the propagation components required for the generation of the current business action result. S226. When the attention head output changes the action consistency judgment result or triggers a sensitive attribute commitment to participate in local attention window aggregation, the corresponding attention head output is written into the privacy inversion message.

[0028] In this invention, changing the action consistency judgment result means that after the attention head output is connected to the local attention window, the semantic state of the business action deviates from the online mall transaction action written in the action query terminal; triggering the participation of sensitive attribute commitment in the local attention window aggregation means that the local attention window calls the propagation component corresponding to the sensitive attribute commitment during the aggregation process; in practice, when any of the above situations occur in the action consistency judgment, the corresponding attention head output is written into the privacy inversion message; the above processing is used to exclude attention head outputs that may change the semantics of the business action or introduce the direction propagation of sensitive attribute commitments from the necessary business messages.

[0029] In this embodiment, the sensitive perturbation sample response detection applied to the privacy inversion message in S3 specifically includes: S31. Freeze the propagation weight of business-necessary messages to the business action result output layer, and close the connection of business-necessary messages to the sensitive disturbance sample response detection loop. In this invention, freezing the propagation weight of essential business messages to the business action result output layer is an operation that fixes the original propagation state of the essential business messages; in practice, the relation-aware attention decomposition operator keeps the propagation weight of the attention head corresponding to the essential business messages unchanged, blocks the essential business messages from entering the sensitive disturbance sample response detection loop, and only allows privacy inversion messages to participate in the sensitive disturbance sample response detection; this operation is used to prevent normal business propagation components from affecting the disturbance response judgment result; In this invention, the sensitive disturbance sample response detection loop is an internal detection path within the model used to detect the impact of privacy inversion messages on business action results. In implementation, the sensitive disturbance sample response detection loop receives privacy inversion messages, keeps necessary business messages from entering the detection path, and limits the scope of subsequent disturbance effects to the attention head and relation edge corresponding to the privacy inversion message. This loop is used to locate the privacy inversion propagation position. S32. Integrate the privacy inversion message into the sensitive disturbance sample response detection loop, and apply paired sensitive disturbance samples with opposite directions and consistent amplitudes within the same attention head; In this invention, paired sensitive perturbation samples are two sets of sensitive perturbation inputs with opposite directions and the same perturbation amplitude. In practice, positive and negative perturbations are added to the sensitive attribute commitment propagation position corresponding to the privacy inversion message, respectively. The two sets of perturbations use the same amplitude but opposite directions. Paired sensitive perturbation samples are used to form a basis for canceling response comparison under the same propagation conditions. In this invention, a detection operation is performed to keep the position of the perturbation effect consistent by applying paired sensitive perturbation samples within the same attention head. In practice, the sensitive perturbation sample response detection loop selects the privacy inversion message propagation position within the same attention head, and applies the forward perturbation and the reverse perturbation to the propagation position respectively, so as to avoid the propagation difference between different attention heads interfering with the output difference judgment. S33. Keep the action query end, business stage and relationship type unchanged, lock the attention head transmission order and relationship edge propagation path of the pair of sensitive disturbance samples, and transmit the pair of sensitive disturbance samples to the business action result output layer respectively. In this invention, locking the attention head transmission order and the relation edge propagation path is an operation to keep the propagation conditions of the two sets of sensitive perturbation samples consistent. During implementation, the action query end, business stage, and relation type remain unchanged, and the paired sensitive perturbation samples are transmitted to the business action result output layer along the same attention head transmission order and the same relation edge propagation path. This operation is used to concentrate the source of output difference on the sensitive perturbation sample itself. In this invention, transmitting paired sensitive perturbation samples to the business action result output layer is a detection operation that generates output differences. During implementation, the privacy inversion messages corresponding to the positive perturbation and the privacy inversion messages corresponding to the negative perturbation enter the business action result output layer through the locked propagation path, and the business action result output layer outputs the corresponding business action result category or output response value, respectively. This output is used for subsequent response feedback. S34. Execute response feedback from the business action result output layer to each attention head, and lock the perturbation sensitive attention head according to the output difference caused by the paired sensitive perturbation samples. In this invention, response backflow is an operation that traces the source of output difference from the business action result output layer in reverse. In practice, the business action result output layer compares the output difference caused by positive and negative perturbations and backflows the output difference to each attention head along the original attention head transmission order. Response backflow is used to determine which attention heads produce a significant response to sensitive perturbation samples. In this invention, the perturbation-sensitive attention head is obtained by response backflow locking; in practice, after the response backflow reaches each attention head, the output difference contribution corresponding to each attention head is compared, and the attention head whose output difference contribution reaches the locking condition is determined as the perturbation-sensitive attention head; the perturbation-sensitive attention head is used as the processing position for subsequent masking relationship edges; S35. Within the perturbation-sensitive attention head, mask each relation edge one by one, compare the output difference before and after masking, and lock the masked relation edge as the perturbation-sensitive relation edge when the output difference decreases to the elimination threshold. In this invention, the step-by-step shielding of relational edges is a detection operation that disconnects the propagation connections of relations one by one within the perturbation-sensitive attention head. In practice, while keeping other relational edges unchanged within the perturbation-sensitive attention head, the propagation connections of individual relational edges are closed one by one, and the output differences before and after closure are compared. This operation is used to find relational edges that cause sensitive perturbation samples to affect the results of business actions. In this invention, the perturbation-sensitive relation edge is obtained by reducing the output difference after each relation edge is masked; in practice, when a relation edge is masked, the output difference corresponding to the business action result output layer decreases to the elimination threshold, and the masked relation edge is locked as a perturbation-sensitive relation edge; the perturbation-sensitive relation edge is used to determine the target of subsequent null projection links; S36. Add and compare the two sets of opposite output differences corresponding to the perturbation-sensitive attention head and the perturbation-sensitive relation edge. When the addition result drops to the cancellation threshold, the perturbation-sensitive attention head and the perturbation-sensitive relation edge are identified as the attention head and relation edge that cause the business action result to change with the sensitive perturbation sample.

[0030] In this invention, the addition and comparison of two sets of opposite output differences is an operation to verify whether the disturbance response comes from the same sensitive propagation direction. In practice, the positive disturbance output difference and the reverse disturbance output difference corresponding to the disturbance sensitive attention head and the disturbance sensitive relation edge are read, and the two sets of output differences are added and compared. If the addition result drops to the cancellation threshold, it means that the two sets of opposite disturbances form a cancellation response on the attention head and relation edge.

[0031] In this invention, the attention head and relation edge that cause the business action result to change with the sensitive disturbance sample are determined by the cancellation response. In practice, only attention heads and relation edges that simultaneously satisfy the following conditions are determined as attention heads and relation edges that cause the business action result to change with the sensitive disturbance sample: response backflow locking, output difference decrease after relation edge shielding, and the sum of the output differences of two sets of opposite directions decreases to the cancellation threshold. This determination result is used for null space projection in S4.

[0032] In this embodiment, step S4 involves performing null-space projection on the attention head and relation edges obtained from the localization, specifically including: S41. Lock the perturbation-sensitive relation edge in the perturbation-sensitive attention head, and import the privacy inversion message corresponding to the perturbation-sensitive relation edge into the null space projection link; In this invention, the privacy inversion message corresponding to the perturbation-sensitive relation edge is the propagation component that has been propagated on the perturbation-sensitive relation edge and written into the privacy inversion message by the relation-aware attention decomposition operator. In practice, based on the perturbation-sensitive attention head and the perturbation-sensitive relation edge locked in S3, the privacy inversion message on the relation edge is read, and the privacy inversion message is transferred from the original attention head output side to the null projection link. The null projection link is used to perform direction stripping processing on the above privacy inversion message. S42. Within the sensitive attention head of the disturbance, disconnect the shared aggregation entry of privacy inversion messages and business-necessary messages, and import the privacy inversion messages separately into the inversion direction stripping entry; In this invention, the shared aggregation entry point is the location where privacy inversion messages and business-necessary messages originally entered the attention aggregation together within the perturbation-sensitive attention head. In practice, the privacy inversion messages are blocked from entering the shared aggregation location within the perturbation-sensitive attention head, and only the aggregation input of business-necessary messages in the original propagation path is retained. The privacy inversion messages are transferred to the inversion direction stripping entry point, so that the privacy inversion messages and business-necessary messages are separated before the null projection is performed. In this invention, the inversion direction stripping entry is the entry point for receiving privacy inversion messages in the null projection link; in practice, the inversion direction stripping entry only receives privacy inversion messages corresponding to perturbation sensitive relation edges, and does not receive business-necessary messages; the inversion direction stripping entry is used to send privacy inversion messages into subsequent orthogonal suppression processing to avoid business-necessary messages being weakened in null projection. S43. Using the propagation direction of the perturbation sensitive relation edge as the null space constraint direction, perform orthogonal suppression on the privacy inversion message in the inversion direction stripping entry; In this invention, the null space constraint direction is obtained from the propagation direction of the perturbation sensitive relation edge; in practice, the propagation direction of the perturbation sensitive relation edge within the perturbation sensitive attention head is read, and this propagation direction is used as the direction reference that needs to be weakened in the privacy inversion message; the null space constraint direction is used to limit the action direction of orthogonal suppression, so that the propagation component pointing to the sensitive attribute commitment direction in the privacy inversion message is suppressed; In this invention, orthogonal suppression is a processing operation that weakens the privacy inversion message according to the null space constraint direction. During implementation, the privacy inversion message in the inversion direction stripping entry is aligned with the null space constraint direction. Propagation components that do not point to the null space constraint direction are retained, while those pointing to the null space constraint direction are suppressed. When performing orthogonal suppression, the propagation weights in the privacy inversion message that are consistent with the null space constraint direction are adjusted to a low-weight state, while the propagation weights that are inconsistent with the null space constraint direction are retained. The privacy inversion message in the low-weight state only participates in subsequent weight merging as a null space projection result and is no longer restored to an independent propagation message. Orthogonal suppression is used to reduce the weight of the privacy inversion message that continues to propagate along the sensitive attribute commitment direction. S44. Feed the privacy inversion message after orthogonal suppression back to the output side of the perturbation sensitive attention head, and retain the propagation weight of business-necessary messages in the perturbation sensitive attention head; In this invention, the operation of reconnecting to the output side of the perturbation-sensitive attention head is to send the privacy inversion message after orthogonal suppression back to the output position of the perturbation-sensitive attention head. In practice, the privacy inversion message after orthogonal suppression no longer enters the shared aggregation entry point, but is reconnected to the output side of the perturbation-sensitive attention head to participate in the weight merging before output. Business-essential messages are still transmitted along the original propagation path, and the propagation weight of business-essential messages is not changed due to the orthogonal suppression of privacy inversion messages. S45. Perform post-projection weight reset on the perturbation-sensitive relation edge and close the attention weight transfer path from the perturbation-sensitive relation edge to the sensitive attribute commitment direction; In this invention, the post-projection weight reset is the operation of attention weight control on the perturbation sensitive relation edge after the null space projection is completed; in practice, after the perturbation sensitive relation edge completes orthogonal suppression, the attention weight state of the relation edge leading to the sensitive attribute commitment direction is reset, and the path for the relation edge to continue to pass attention weight to the sensitive attribute commitment direction is closed; the post-projection weight reset is used to prevent the privacy inversion message after orthogonal suppression from re-propagating in the sensitive attribute commitment direction; S46. Before the perturbation-sensitive attention head completes its output, the propagation weights of the business-necessary messages and the privacy-inverting messages after orthogonal suppression are weighted and merged. The original propagation path of the business-necessary messages is retained, and the privacy-inverting messages after orthogonal suppression are limited to null-space projection results to participate in the weight merging.

[0033] In this invention, weight merging is an operation performed on the output side to merge business-essential messages and orthogonally suppressed privacy-inverted messages before the perturbation-sensitive attention head completes its output. In practice, business-essential messages participate in the output according to their original propagation path and original propagation weights, while orthogonally suppressed privacy-inverted messages only participate in the output-side merging in the form of null projection results and are no longer restored to independently transmittable privacy-inverted messages. Weight merging is used to preserve the supporting role of business-essential messages in the results of business actions and to restrict privacy-inverted messages to a controlled state after projection.

[0034] In this embodiment, the disturbance invariance gating result obtained in S5 specifically includes: S51. Inject the necessary business messages after zero-space projection back into the action query terminal, and lock the action query terminal, business stage, relationship type and attention head transmission order. In this invention, the back-injection action query terminal is the operation of sending back the necessary business messages that still retain their business propagation function after null space projection to the action query terminal. In practice, the necessary business messages after null space projection in S4 are read and sent back to the corresponding business action query position of the action query terminal, keeping the order of action query terminal, business stage, relationship type and attention head transmission unchanged. The above operation is used to provide fixed propagation conditions for disturbance invariance gating and avoid the impact of changes in propagation order on the two class comparisons. In this invention, the attention head transmission order is obtained by the attention head propagation path that remains consistent before and after null space projection; during implementation, the order of attention heads that the business-necessary messages pass through is locked according to the original propagation path of the business-necessary messages retained in S4; the attention head transmission order is used to limit the propagation conditions when closing sensitive disturbance samples and opening sensitive disturbance samples, so that the result categories of the two business actions are comparable. S52. Within the disturbance invariance gating, sensitive disturbance samples are closed and opened respectively, and the gating baseline category and business action result category are obtained according to the locked attention head transmission order. In this invention, closing sensitive disturbance samples means temporarily not applying sensitive disturbance samples within the disturbance invariance gating. In practice, the necessary business messages after injection enter the business action result output layer according to the locked attention head transmission order, and the category output by the business action result output layer is used as the gating benchmark category. The gating benchmark category is used as the category comparison benchmark after the sensitive disturbance samples are subsequently enabled. In this invention, enabling sensitive disturbance samples is an operation that applies sensitive disturbance samples under the same propagation conditions. During implementation, the order of action query end, business stage, relationship type and attention head transmission remains unchanged. The necessary business messages after back-injection re-enter the business action result output layer under the sensitive disturbance sample enabling state. The category output by the business action result output layer is used as the business action result category. The business action result category is used to perform category flipping judgment with the gating benchmark category. S53. When the business action result category is consistent with the gating benchmark category, the output connection of the business action result output layer is retained. In this invention, the consistency between the business action result category and the gating benchmark category indicates that the category judgment of the business action result output layer has not been changed after the sensitive disturbance sample is enabled. In practice, the disturbance invariance gating compares the business action result category and the gating benchmark category. When they are consistent, the output connection of the business action result output layer is kept in a connected state. The above processing is used to allow business action results that are not affected by sensitive disturbance samples to continue to enter the subsequent inversion budget control process. S54. When the business action result category is flipped relative to the gating baseline category, the output connection of the business action result output layer is cut off, and the corresponding business action result output is blocked. In this invention, category flipping indicates that the category of the business action result obtained after the sensitive disturbance sample is enabled is inconsistent with the gating baseline category. In practice, after the disturbance invariance gating detects category flipping, it cuts off the output connection of the business action result output layer, so that the corresponding business action result is no longer passed to the subsequent process. The above processing is used to block the output of business action results that change with the sensitive disturbance sample. S55. Write the corresponding states in the output connection retention state and output connection disconnection state of the business action result output layer into the disturbance invariant gating result.

[0035] In this invention, the disturbance invariance gating result is obtained after the disturbance invariance gating completes the category comparison. During implementation, when the category of the business action result is consistent with the gating benchmark category, the output connection retention state of the business action result output layer is written into the disturbance invariance gating result. When the category of the business action result is flipped relative to the gating benchmark category, the output connection disconnection state of the business action result output layer is written into the disturbance invariance gating result. The disturbance invariance gating result is used for subsequent inversion budget consumption value accumulation and smart contract shadow replay verification.

[0036] In this embodiment, the cumulative inversion budget consumption value in S6 specifically includes: S61. Write the retained business action result categories in reverse to the sensitive perturbation samples corresponding to the perturbation sensitive attention head and the perturbation sensitive relation edge, and mark the sensitive perturbation samples excluded by the business action result categories. In this invention, the retained business action result category is the business action result category obtained after connecting the output of the disturbance-invariant gated retained business action result output layer; in implementation, the retained business action result category is written back to the corresponding sensitive disturbance sample along the disturbance-sensitive attention head and disturbance-sensitive relation edge, and the retainable state of the sensitive disturbance sample before and after the write-back is compared; the sensitive disturbance samples excluded by the business action result category are marked as samples participating in the inversion budget accumulation; In this invention, reverse writing is the operation of sending the business action result category back to the propagation location of the sensitive disturbance sample. In practice, reverse writing does not change the action query end, business stage, and relationship type, but only applies the retained business action result category to the sensitive disturbance sample corresponding to the disturbance sensitive attention head and the disturbance sensitive relationship edge. Reverse writing is used to detect the ability of the business action result category to exclude sensitive disturbance samples. S62. Accumulate and invert the budget consumption value by taking the marked sensitive disturbance sample along the budget back pressure gate and the corresponding disturbance sensitive attention head and disturbance sensitive relation edge. In this invention, the budget backpressure gate is a gating operation that accumulates the inversion budget consumption value along the disturbance-sensitive attention head and the disturbance-sensitive relationship edge. During implementation, the budget backpressure gate reads the marked sensitive disturbance samples, backpressures them along the corresponding disturbance-sensitive attention head and disturbance-sensitive relationship edge to the business action result output layer, and accumulates the inversion budget consumption value according to the proportion of sensitive disturbance samples excluded by the business action result category to the total number of corresponding sensitive disturbance samples. After each business action result category excludes sensitive disturbance samples, the budget backpressure gate adds the current exclusion proportion to the previous round of inversion budget consumption value to obtain the updated inversion budget consumption value. When there is no previous round of inversion budget consumption value, the current exclusion proportion is used as the initial inversion budget consumption value. In this invention, the inversion budget consumption value represents the cumulative value of the business action result category's ability to exclude sensitive disturbance samples; during implementation, the inversion budget consumption value increases as the proportion of excluded sensitive disturbance samples increases; the inversion budget consumption value is used to determine whether the current business action result category has an excessively strong inversion capability, serving as the basis for subsequent control measures such as maintaining output, reducing granularity, or output circuit breaking. S63. When the inversion budget consumption value does not reach the threshold, maintain the current output connection of the business action result output layer; In this invention, the threshold is used to determine whether the inverted budget consumption value exceeds the allowable output range. In practice, the inverted budget consumption value is a cumulative value between 0 and 1, and the threshold is 0.70. When the inverted budget consumption value is less than 0.70, the budget back pressure gate does not change the current output connection of the business action result output layer, and the business action result category continues to maintain the original granularity output. S64. When the inversion budget consumption value reaches the threshold, the business action result category is converted into the higher-level granular business action result category. In this invention, the higher-level granularity business action result category is a coarser output category than the business action result category. During implementation, when the inversion budget consumption value reaches 0.70, the original business action result category is converted from a fine-grained category to a higher-level granularity business action result category to reduce the ability of the business action result category to exclude sensitive perturbation samples. The higher-level granularity business action result category is used to reduce privacy inversion risks while preserving the meaning of business processing. S65. After the business action result category is converted, if the budget back pressure gate accumulates and inverts the budget consumption value again and still reaches the threshold, the output fuse gate cuts off all output connections of the business action result output layer and terminates the business action result output.

[0037] In this invention, the budget consumption value is accumulated again after the business action result category is converted to the higher-level granularity business action result category and the budget backpressure operation is re-executed. During implementation, the budget backpressure gate accumulates the budget consumption value again along the perturbation sensitive attention head and perturbation sensitive relationship edge corresponding to the marked sensitive perturbation sample. When the accumulated result still reaches 0.70, the output circuit breaker cuts off all output connections of the business action result output layer, and the business action result no longer enters the external output terminal.

[0038] In this invention, the output circuit breaker is a gating operation that terminates the output when the inversion budget control requirements cannot be met even after reducing the granularity. In practice, the output circuit breaker receives the inversion budget consumption value accumulated again by the budget back pressure gate. When the result of the re-accumulation still reaches the threshold, all output connections of the business action result output layer are closed. The output circuit breaker is used to block the output of business action results that still have privacy inversion risks.

[0039] In this embodiment, the release verification is performed through a smart contract in S7, specifically including: S71. Cut off the direct connection between the business action result output layer and the external output end within the chain certificate release layer, and connect the business action result output layer to the smart contract shadow replay channel. In this invention, the chain certificate release layer is an on-chain verification and control layer set between the business action result output layer and the external output end. In implementation, the chain certificate release layer receives the output request from the business action result output layer, first closes the direct connection channel between the business action result output layer and the external output end, and then connects the business action result output layer to the smart contract shadow replay channel. The chain certificate release layer is used to block the direct output of business action results that have not been verified on the chain. In this invention, the direct connection channel from the business action result output layer to the external output terminal is the path through which the business action result output layer bypasses the smart contract release verification and outputs directly to the outside. In practice, the chain certificate release layer sets the above-mentioned direct connection channel to a disconnected state, and the output request of the business action result output layer can only enter the smart contract shadow replay channel. This process is used to ensure that the business action result must be reviewed on-chain before being output to the outside. S72. Write the null space projection results, perturbation invariance gating results, and inversion budget consumption value into the blockchain; In this invention, writing the null projection result, the perturbation invariance gating result, and the inversion budget consumption value into the blockchain is an on-chain record operation for fixing the internal control results of the three types of models. During implementation, the blockchain certificate release layer reads the null projection result formed by S4, the perturbation invariance gating result formed by S5, and the inversion budget consumption value formed by S6, and writes the three types of results into the blockchain. The blockchain record is used for reading and verification by the smart contract shadow replay channel to prevent the release verification from deviating from the control results of the preceding model. S73, The smart contract shadow replay channel performs shadow replay on the null space projection result, the disturbance invariance gating result and the inversion budget consumption value to generate the first release flag, the second release flag and the third release flag; In this invention, the smart contract shadow replay channel is an on-chain replay path that performs verification based on the internal control results of three types of models in the blockchain. During implementation, the smart contract shadow replay channel reads the null projection result, the perturbation invariance gating result, and the inversion budget consumption value. It does not retrain the heterogeneous graph Transformer, but only verifies whether the control actions of the preceding model remain effective according to the on-chain records. After the verification is passed, the first release flag, the second release flag, and the third release flag are generated respectively. In this invention, the first release flag, the second release flag, and the third release flag are on-chain release signals written to the release latch after the smart contract shadow replay channel completes three types of verification. In practice, the first release flag corresponds to the verification pass status of the null projection result, the second release flag corresponds to the verification pass status of the perturbation invariance gating result, and the third release flag corresponds to the verification pass status of the inversion budget consumption value. The three types of release flags are used to jointly determine whether the business action result output layer is connected to the external output terminal. S74. After the first release flag, the second release flag, and the third release flag are all written, the release latch connects the business action result output layer with the external output terminal. In this invention, the release latch is a structure that controls the connection between the output layer of the business action result in the chain release layer and the external output terminal. In practice, the release latch reads the write status of the first release mark, the second release mark and the third release mark. After all three types of release marks are written, the release latch connects the business action result output layer with the external output terminal, and the business action result enters the external output terminal. In this invention, all three types of release flags are written to indicate that the null projection result, the perturbation invariance gating result, and the inversion budget consumption value are all verified through the smart contract shadow replay channel. In practice, the release latch does not perform individual release for any single release flag, but only opens the output connection when the first release flag, the second release flag, and the third release flag all exist. This process is used to avoid directly releasing the business action result after a single path has passed the verification. S75. If any release flag is not written, the release latch keeps the business action result output layer disconnected from the external output terminal and writes the shadow replay result corresponding to the missing release flag into the abnormal audit record.

[0040] In this invention, if any release flag is not written, it indicates that at least one of the following has failed the review by the smart contract shadow replay channel: null projection result, disturbance invariance gating result, or inversion budget consumption value. In practice, after the release latch detects that any one of the first, second, and third release flags is missing, it keeps the business action result output layer disconnected from the external output terminal, and the business action result is not output to the outside.

[0041] In this invention, the shadow replay result corresponding to the missing release marker is the review result left when the smart contract shadow replay channel fails to generate the corresponding release marker. In practice, the smart contract shadow replay channel reads the shadow replay result corresponding to the missing release marker and writes the missing release marker type, business stage, business action result category and review failure reason into the abnormal audit record. The abnormal audit record is written into the blockchain for subsequent tracing of the reason why the business action result was not released.

[0042] In this embodiment, the shadow replay is performed in S73, specifically including: S731, the smart contract shadow replay channel includes the projection weight reset verification path, the category flip verification path, and the budget circuit breaker verification path; In this invention, the post-projection weight reset verification path, category flip verification path, and budget circuit breaker verification path are obtained by the smart contract shadow replay channel according to the source actions of the three types of on-chain results. During implementation, the smart contract shadow replay channel reads the null projection result, the perturbation invariance gating result, and the inverted budget consumption value. It maps the null projection result to the post-projection weight reset action, the perturbation invariance gating result to the retained output connection action, and the inverted budget consumption value to the budget backpressure gate and output circuit breaker gate actions. According to the above three correspondences, the post-projection weight reset verification path, the category flip verification path, and the budget circuit breaker verification path are formed in the smart contract shadow replay channel. The three verification paths are used to verify whether the null projection result, the perturbation invariance gating result, and the inverted budget consumption value meet the release conditions. S732, read the null space projection result of the post-projection weight reset verification path, replay the post-projection weight reset action, and write the first release flag when the attention weight transfer path from the perturbation sensitive relation edge to the sensitive attribute commitment direction remains closed. In this invention, the post-projection weight reset verification path is the on-chain verification path that verifies the weight reset state corresponding to the null projection result. During implementation, the post-projection weight reset verification path reads the perturbation-sensitive relation edge, the post-projection weight reset state, and the attention weight transfer path state from the null projection result, and replays the post-projection weight reset action according to the null projection result. If the attention weight transfer path from the perturbation-sensitive relation edge to the sensitive attribute commitment direction remains closed, the post-projection weight reset verification path writes a first pass flag. The first pass flag indicates that the post-projection weight reset action corresponding to the null projection result has passed on-chain verification. S733, the category flip verification path reads the disturbance invariance gating result, replays the retained output connection action, and writes the second release flag when the business action result category has not been flipped relative to the gating base category. In this invention, the category flip verification path is an on-chain verification path that verifies the stable state of the category corresponding to the perturbation invariance gating result. During implementation, the category flip verification path reads the gating baseline category, business action result category, and output connection retention status from the perturbation invariance gating result, and replays the retained output connection action according to the perturbation invariance gating result. If the business action result category has not been flipped relative to the gating baseline category, and the output connection retention status is consistent with the category comparison result, then the category flip verification path writes a second release flag. The second release flag indicates that the retained output connection action corresponding to the perturbation invariance gating result has passed on-chain verification. S734, the budget circuit breaker verification path reads the inverted budget consumption value, replays the budget back pressure gate and output circuit breaker gate actions, and writes the third release flag when the output circuit breaker gate is not triggered.

[0043] In this invention, the budget circuit breaker verification path is an on-chain verification path that verifies the output allowable state corresponding to the inverted budget consumption value. During implementation, the budget circuit breaker verification path reads the inverted budget consumption value, the budget back pressure gate state, and the output circuit breaker gate state, and replays the budget back pressure gate and output circuit breaker gate actions according to the inverted budget consumption value. If the output circuit breaker gate is not triggered, the budget circuit breaker verification path writes a third release flag. The third release flag is used to indicate that the budget control action corresponding to the inverted budget consumption value has passed the on-chain verification.

[0044] In this embodiment, the release latch in S74 includes a first release tag latch bit, a second release tag latch bit, a third release tag latch bit, and an output connection control bit. When all three release tag latch bits are in the write state, the output connection control bit opens the output connection between the service action result output layer and the external output terminal. When any release tag latch bit is not written, the output connection control bit closes the output connection between the service action result output layer and the external output terminal. After the service action result is output, the three release tag latch bits are cleared.

[0045] In this invention, the release latch is a structure within the chain certificate release layer that controls the output connection based on the first release flag, the second release flag, and the third release flag. In implementation, the release latch receives the first release flag, the second release flag, and the third release flag written by the smart contract shadow replay channel, and does not directly read the null space projection result, the perturbation invariance gating result, and the inversion budget consumption value. The release latch is used to connect the business action result output layer with the external output end after the verification on the three types of chains has passed.

[0046] In this invention, the first release flag latch, the second release flag latch, and the third release flag latch are obtained from the write positions of the first release flag, the second release flag, and the third release flag. In practice, the first release flag is written to the first release flag latch, the second release flag is written to the second release flag latch, and the third release flag is written to the third release flag latch. The three release flag latches are used to store the pass status of the three verification paths respectively.

[0047] In this invention, the output connection control bit is the control position for the output connection status between the control service action result output layer inside the release latch and the external output terminal. In practice, the output connection control bit reads the write status of the three release mark latch bits. When the first release mark latch bit, the second release mark latch bit, and the third release mark latch bit are all in the write state, the output connection control bit opens the output connection between the service action result output layer and the external output terminal.

[0048] In this invention, the absence of any release tag latch bit indicates that at least one type of release tag among the first release tag, the second release tag, and the third release tag has not been written through the corresponding verification path. In practice, when the output connection control bit detects that any release tag latch bit has not been written, it closes the output connection between the service action result output layer and the external output terminal, and the service action result does not enter the external output terminal. The above control is used to prevent the service action result from being released prematurely after a single release tag is written.

[0049] In this invention, clearing the three release flag latch bits is a reset operation after the business action result is output. In practice, after the business action result is output through the external output terminal, the release latch clears the first release flag latch bit, the second release flag latch bit, and the third release flag latch bit, and restores the output connection control bit to the closed state. Before the next business action result is output, the release latch waits again for the first release flag, the second release flag, and the third release flag to be written. Before the three types of release flags are rewritten, the output connection control bit remains in the closed state, and the release state formed when the previous business action result was output is not used. The above reset operation is used to prevent the release flag formed by the previous release verification from being used by the subsequent business action result.

[0050] Example 1: To verify the feasibility of this invention in practice, it was applied to a transaction review and access approval scenario in an online shopping mall. In this scenario, external access requests originate from user order placement, payment verification, refund review, after-sales verification, and audit query stages. Traditional methods primarily rely on account permissions, field anonymization, and blockchain records, which can limit the leakage of sensitive fields but struggle to determine whether the business action results themselves carry clues related to sensitive attributes. The platform frequently needs to output business action results such as approval, review, delayed processing, and rejection during refund review and after-sales verification. If the model incorporates propagation components such as sensitive attribute commitments, address associations, and risk tags into the attention head output when generating results, external access parties, even if they cannot see sensitive fields, can infer some sensitive attributes through category changes, fine-grained results, and differences across multiple accesses.

[0051] In this embodiment, online mall transaction actions include order placement, payment verification, refund review, after-sales verification, and audit verification; business stages include transaction creation, payment confirmation, logistics flow, refund processing, after-sales verification, and audit verification; sensitive attribute commitments are obtained by commitment-encoded user identity attributes, transaction preference attributes, account risk attributes, and address association attributes; contract verification relationships are obtained by the correspondence between on-chain release conditions, model output status, and audit records. After an external access request enters the system, it does not directly trigger the reading of sensitive fields, but instead writes to the action query end of the heterogeneous graph Transformer. The action query end initiates multi-head attention propagation around the business action query.

[0052] After receiving heterogeneous messages, the action query terminal connects to the output side of each attention head via a relation-aware attention decomposition operator. Transaction action relationships, business stage relationships, sensitive attribute commitment relationships, and contract verification relationships are propagated through their respective attention heads to form heterogeneous messages. The relation-aware attention decomposition operator performs relation projection, business stage masking, and action consistency judgment on the heterogeneous messages. Attention head outputs that pass through the business stage mask without changing the action consistency judgment result are written into the business-necessary message. Attention head outputs that deviate from the current business stage, change the semantics of the business action, or trigger sensitive attribute commitments to participate in local attention window aggregation are written into the privacy inversion message. The business-necessary message retains the propagation connection to the business action result output layer, while the privacy inversion message is transferred to the sensitive perturbation sample response detection loop.

[0053] In the privacy-reversal message detection stage, the system freezes the propagation weights of essential business messages to the business action result output layer and closes the connection for essential business messages to enter the sensitive disturbance sample response detection loop. The detection loop applies pairs of sensitive disturbance samples with opposite directions and consistent amplitudes within the same attention head, keeping the action query end, business stage, and relationship type unchanged. The two sets of sensitive disturbance samples are sent to the business action result output layer along the same attention head propagation order and the same relation edge propagation path. After the business action result output layer generates an output difference, the difference flows back along the attention head, and the system locks the disturbed sensitive attention head. Within the disturbed sensitive attention head, relation edges are blocked one by one. When the output difference decreases to the elimination threshold, the blocked relation edges are locked as disturbed sensitive relation edges.

[0054] Once the perturbation-sensitive attention heads and perturbation-sensitive relation edges are determined, the system imports the corresponding privacy-reversing messages into the null-space projection link. The null-space projection link uses the propagation direction of the perturbation-sensitive relation edges as the null-space constraint direction, performs orthogonal suppression on the privacy-reversing messages, resets the weights of the perturbation-sensitive relation edges after projection, and closes the attention weight propagation path from the perturbation-sensitive relation edges to the sensitive attribute commitment direction. The orthogonally suppressed privacy-reversing messages only participate in weight merging in the form of null-space projection results; business-essential messages retain their original propagation paths and propagation weights. This processing enables the model to perform targeted weakening of the privacy-reversing propagation path before outputting business action results, rather than filtering after the results are output.

[0055] After completing null-space projection, the necessary business messages are injected back into the action query terminal, and the action query terminal, business stage, relationship type, and attention head transmission order remain unchanged. The perturbation invariance gating is used to close and open sensitive perturbation samples respectively, obtaining the gating baseline category and business action result category according to the same attention head transmission order. When the two results are consistent, the output connection of the business action result output layer is retained; when category flipping occurs, the output connection is severed. For the retained business action result categories, the budget backpressure gate accumulates the inverted budget consumption value along the perturbation-sensitive attention head and perturbation-sensitive relationship edge. When the inverted budget consumption value is less than 0.70, the current granularity output is maintained; when it reaches 0.70, the business action result category is converted to a higher-level granularity business action result category; if it still reaches 0.70 after conversion, the output circuit breaker gate severs all output connections of the business action result output layer.

[0056] The blockchain-based certificate release layer is positioned between the business action result output layer and the external output terminal. Before the business action result is output externally, the blockchain-based certificate release layer cuts off the direct connection between the business action result output layer and the external output terminal, writing the zero-space projection result, the perturbation invariance gating result, and the inverted budget consumption value into the blockchain. The smart contract shadow replay channel reads the three types of results and performs verification through the projection weight reset verification path, the category flip verification path, and the budget circuit breaker verification path, respectively. The three verification paths correspond to the first release flag, the second release flag, and the third release flag. The release latch connects the business action result output layer and the external output terminal only after all three release flags have been written; if any release flag is missing, the business action result is not output externally, and the shadow replay result corresponding to the missing release flag is written into the anomaly audit record.

[0057] To verify the feasibility of this invention in practice, it was applied to the risk control and transaction review scenario of an online shopping mall. A comparison was made between the conventional field anonymization and on-chain evidence storage method and the method of this invention. The verification sample included 120,000 external access requests, covering order placement, payment verification, refund review, after-sales verification, and audit verification. During verification, the conventional method retained field anonymization, permission verification, and on-chain records; the method of this invention added relation-aware attention decomposition, sensitive perturbation sample response detection, null projection, perturbation invariance gating, inversion budget control, and smart contract shadow replay to the same request entry point. Statistical results show that this invention significantly improves the privacy inversion hit rate, category flip rate, accuracy of sensitive relation edge identification, and anomaly audit completeness rate, with only a slight increase in business response latency and no decrease in the retention rate of normal transaction actions.

[0058] Table 1: Comparison of the Implementation Effects of Data Protection in Online Shopping Malls

[0059] The data in Table 1 shows that conventional methods can achieve field-level protection and on-chain records, but they are insufficient in controlling the inversion clues in the results of business actions. The privacy inversion hit rate decreased from 14.8% to 4.1%, and the class flip rate under sensitive perturbations decreased from 11.6% to 2.3%, indicating that relation-aware attention decomposition and null space projection can separate the propagation components of sensitive attribute commitment direction from the necessary business propagation, and the results of business actions no longer fluctuate significantly with sensitive perturbation samples.

[0060] The accuracy rate of sensitive relationship edge identification improved from 71.5% to 89.3%, reflecting that the combination of response backflow and line-by-line blocking of relationship edges can more accurately locate the privacy inversion propagation path. The retention rate of normal transaction actions improved from 92.4% to 96.8%, indicating that this invention does not rely on simply denying access to achieve privacy protection, but rather weakens privacy inversion messages while retaining the propagation weight of necessary business messages. The proportion of high-risk fine-grained output decreased by 43.8 percentage points, indicating that inversion budget control can reduce the granularity of business action results when risks increase.

[0061] The on-chain release mechanism has yielded significant improvements: the number of outputs without on-chain verification has decreased from 37 per 10,000 to 0, and the anomaly audit record completeness rate has increased from 81.2% to 99.1%. The smart contract shadow replay channel records not only the authorization status but also the weight reset after verification projection, maintaining output connections, budget backpressure gates, and output circuit breaker gates. While the average business response latency has increased slightly from 186 milliseconds to 203 milliseconds, this small increase has resulted in a significant improvement in the effectiveness of circuit breaker triggering after budget overruns and the anomaly audit completeness rate, meeting the comprehensive requirements of online marketplace audit scenarios for both security and processing efficiency.

[0062] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A data protection method for a blockchain-based online marketplace system, characterized in that, Includes the following steps: S1. Embed online mall transaction actions, business stages, sensitive attribute commitments and contract verification relationships into the action query end of the heterogeneous graph Transformer, and write external access requests into the action query end. S2. Using the relation-aware attention decomposition operator, the heterogeneous messages received by the action query terminal are decomposed into business-necessary messages and privacy-reversal messages. S3. Apply sensitive perturbation sample response detection to privacy inversion messages to locate attention heads and relation edges that cause changes in business action results with sensitive perturbation samples; S4. Perform null-spatial projection on the attention heads and relation edges obtained from the positioning to reduce the propagation weight of privacy inversion messages and retain the propagation weight of business-necessary messages. S5. Inject the necessary business messages after null space projection back into the action query terminal, perform perturbation invariance gating on the business action results, and obtain perturbation invariance gating results, retaining only the business action results that do not undergo class flipping with sensitive perturbation samples. S6. Accumulate the inversion budget consumption value according to the exclusion ratio of sensitive disturbance samples based on the retained business action results. When the inversion budget consumption value reaches the threshold, reduce the granularity of the business action results. If the threshold is still reached after reducing the granularity, terminate the output of the business action results. S7. Write the null space projection result, perturbation invariance gating result, and inversion budget consumption value into the blockchain, execute the release verification through smart contract, output the business action result after the verification passes, and write the abnormal audit record when the verification fails.

2. The data protection method for a blockchain-based online marketplace system according to claim 1, characterized in that, S2 specifically includes: S21. Disconnect the multi-head attention aggregation link between the action query end and the business action result output layer in the heterogeneous graph Transformer, and cut off the connection that heterogeneous messages directly enter the business action result output layer after being weighted by attention heads. S22. In each attention head, a relation-aware attention decomposition operator is connected to perform relation projection, business stage masking, and action consistency judgment on each heterogeneous message in sequence. S23. When the heterogeneous message after relation projection passes through the business stage mask and is judged by action consistency, the output of the corresponding attention head is written into the necessary business message. S24. When the heterogeneous message after relation projection fails any judgment in the business stage mask or action consistency judgment, write the output of the corresponding attention head into the privacy inversion message. S25. Retain the propagation connection to the business action result output layer for business-necessary messages, and close the propagation connection to the business action result output layer for privacy inversion messages. S26. Use business-necessary messages and privacy-inverting messages as the decomposition output of the relation-aware attention decomposition operator.

3. The data protection method for a blockchain-based online marketplace system according to claim 2, characterized in that, The consistency determination of service stage mask and action in S22 specifically includes: S221. Input the relation projection result into the business stage mask, and control the attention head propagation connection according to the current business stage. S222. When the relation projection result falls within the processing scope of the current business stage, open the propagation connection from the corresponding attention head to the action consistency judgment. S223. When the relation projection result is outside the processing scope of the current business stage, the corresponding attention head will be output bypassed to the privacy inversion message. S224. After the attention head output is enabled by the business stage mask, it is connected to the local attention window of the action query terminal, and the semantic offset comparison before and after the connection is performed. S225. When the attention head output does not change the action consistency judgment result and does not trigger the sensitive attribute commitment to participate in the local attention window aggregation, the corresponding attention head output will be written into the business necessary message. S226. When the attention head output changes the action consistency judgment result or triggers a sensitive attribute commitment to participate in local attention window aggregation, the corresponding attention head output is written into the privacy inversion message.

4. The data protection method for a blockchain-based online marketplace system according to claim 2, characterized in that, The sensitive perturbation sample response detection applied to the privacy inversion message in S3 specifically includes: S31. Freeze the propagation weight of business-necessary messages to the business action result output layer, and close the connection of business-necessary messages to the sensitive disturbance sample response detection loop. S32. Integrate the privacy inversion message into the sensitive disturbance sample response detection loop, and apply paired sensitive disturbance samples with opposite directions and consistent amplitudes within the same attention head; S33. Keep the action query end, business stage and relationship type unchanged, lock the attention head transmission order and relationship edge propagation path of the pair of sensitive disturbance samples, and transmit the pair of sensitive disturbance samples to the business action result output layer respectively. S34. Execute response feedback from the business action result output layer to each attention head, and lock the perturbation sensitive attention head according to the output difference caused by the paired sensitive perturbation samples. S35. Within the perturbation-sensitive attention head, mask each relation edge one by one, compare the output difference before and after masking, and lock the masked relation edge as the perturbation-sensitive relation edge when the output difference decreases to the elimination threshold. S36. Add and compare the two sets of opposite output differences corresponding to the perturbation-sensitive attention head and the perturbation-sensitive relation edge. When the addition result drops to the cancellation threshold, the perturbation-sensitive attention head and the perturbation-sensitive relation edge are identified as the attention head and relation edge that cause the business action result to change with the sensitive perturbation sample.

5. The data protection method for a blockchain-based online marketplace system according to claim 4, characterized in that, In step S4, null-space projection is performed on the attention head and relation edges obtained from the localization, specifically including: S41. Lock the perturbation-sensitive relation edge in the perturbation-sensitive attention head, and import the privacy inversion message corresponding to the perturbation-sensitive relation edge into the null space projection link; S42. Within the sensitive attention head of the disturbance, disconnect the shared aggregation entry of privacy inversion messages and business-necessary messages, and import the privacy inversion messages separately into the inversion direction stripping entry; S43. Using the propagation direction of the perturbation sensitive relation edge as the null space constraint direction, perform orthogonal suppression on the privacy inversion message in the inversion direction stripping entry; S44. Feed the privacy inversion message after orthogonal suppression back to the output side of the perturbation sensitive attention head, and retain the propagation weight of business-necessary messages in the perturbation sensitive attention head; S45. Perform post-projection weight reset on the perturbation-sensitive relation edge and close the attention weight transfer path from the perturbation-sensitive relation edge to the sensitive attribute commitment direction; S46. Before the perturbation-sensitive attention head completes its output, the propagation weights of the business-necessary messages and the privacy-inverting messages after orthogonal suppression are weighted and merged. The original propagation path of the business-necessary messages is retained, and the privacy-inverting messages after orthogonal suppression are limited to null-space projection results to participate in the weight merging.

6. The data protection method for a blockchain-based online marketplace system according to claim 5, characterized in that, The disturbance invariance gating result obtained in S5 specifically includes: S51. Inject the necessary business messages after zero-space projection back into the action query terminal, and lock the action query terminal, business stage, relationship type and attention head transmission order. S52. Within the disturbance invariance gating, sensitive disturbance samples are closed and opened respectively, and the gating baseline category and business action result category are obtained according to the locked attention head transmission order. S53. When the business action result category is consistent with the gating benchmark category, the output connection of the business action result output layer is retained. S54. When the business action result category is flipped relative to the gating baseline category, the output connection of the business action result output layer is cut off, and the corresponding business action result output is blocked. S55. Write the corresponding states in the output connection retention state and output connection disconnection state of the business action result output layer into the disturbance invariant gating result.

7. A data protection method for a blockchain-based online marketplace system according to claim 6, characterized in that, The cumulative inversion budget consumption value in S6 specifically includes: S61. Write the retained business action result categories in reverse to the sensitive perturbation samples corresponding to the perturbation sensitive attention head and the perturbation sensitive relation edge, and mark the sensitive perturbation samples excluded by the business action result categories. S62. Accumulate and invert the budget consumption value by taking the marked sensitive disturbance sample along the budget back pressure gate and the corresponding disturbance sensitive attention head and disturbance sensitive relation edge. S63. When the inversion budget consumption value does not reach the threshold, maintain the current output connection of the business action result output layer; S64. When the inversion budget consumption value reaches the threshold, the business action result category is converted into the higher-level granular business action result category. S65. After the business action result category is converted, if the budget back pressure gate accumulates and inverts the budget consumption value again and still reaches the threshold, the output fuse gate cuts off all output connections of the business action result output layer and terminates the business action result output.

8. The data protection method for a blockchain-based online marketplace system according to claim 7, characterized in that, The release verification performed via smart contract in S7 specifically includes: S71. Cut off the direct connection between the business action result output layer and the external output end within the chain certificate release layer, and connect the business action result output layer to the smart contract shadow replay channel. S72. Write the null space projection results, perturbation invariance gating results, and inversion budget consumption value into the blockchain; S73, The smart contract shadow replay channel performs shadow replay on the null space projection result, the disturbance invariance gating result and the inversion budget consumption value to generate the first release flag, the second release flag and the third release flag; S74. After the first release flag, the second release flag, and the third release flag are all written, the release latch connects the business action result output layer with the external output terminal. S75. If any release flag is not written, the release latch keeps the business action result output layer disconnected from the external output terminal and writes the shadow replay result corresponding to the missing release flag into the abnormal audit record.

9. A data protection method for a blockchain-based online marketplace system according to claim 8, characterized in that, The shadow replay operation in S73 specifically includes: S731, the smart contract shadow replay channel includes the projection weight reset verification path, the category flip verification path, and the budget circuit breaker verification path; S732, read the null space projection result of the post-projection weight reset verification path, replay the post-projection weight reset action, and write the first release flag when the attention weight transfer path from the perturbation sensitive relation edge to the sensitive attribute commitment direction remains closed. S733, the category flip verification path reads the disturbance invariance gating result, replays the retained output connection action, and writes the second release flag when the business action result category has not been flipped relative to the gating base category. S734, the budget circuit breaker verification path reads the inverted budget consumption value, replays the budget back pressure gate and output circuit breaker gate actions, and writes the third release flag when the output circuit breaker gate is not triggered.

10. A data protection method for a blockchain-based online marketplace system according to claim 8, characterized in that, The release latch in S74 includes a first release tag latch bit, a second release tag latch bit, a third release tag latch bit, and an output connection control bit. When all three release tag latch bits are in the write state, the output connection control bit opens the output connection between the service action result output layer and the external output terminal. When any release tag latch bit is not written, the output connection control bit closes the output connection between the service action result output layer and the external output terminal. After the service action result is output, the three release tag latch bits are cleared.

Images