Cloud data semantic and integrity double-layer auditing method and system based on hypergraph and zero-knowledge proof

Abstract

The application discloses a cloud data semantic and integrity double-layer auditing method and system based on hypergraph and zero-knowledge proof, and relates to the cross field of cloud computing, data integrity verification and cryptography. The method relates to the zero-knowledge data possession integrity proof of ciphertext, the multi-dimensional hypergraph topology feature extraction, the semantic authentication data structure construction and the double-layer integrity auditing mechanism for realizing the data integrity and the semantic. The method proposes a double-layer integrity auditing architecture which is decoupled from the physical possession and the high-order semantic, realizes the responsibility isolation and the accurate tracing of the bottom physical attack and the high-level business poisoning under the legal identity, proposes a deterministic cryptographic school verification closed loop based on fixed-point quantization and local forward recalculation, proposes a double threshold anti-chronic poisoning mechanism based on absolute reference and single relative offset, and improves the stability of the cloud data semantic auditing process.

Description

Technical Field

[0001] This application relates to the interdisciplinary fields of cloud computing, data integrity verification, and cryptography, and in particular to a cloud data semantic and integrity two-layer auditing method and system based on hypergraphs and zero-knowledge proofs. Background Technology

[0002] As core business data of government and enterprises migrates comprehensively to the cloud, ensuring the integrity and authenticity of outsourced data in untrusted cloud environments has become a core requirement in the field of data security. Existing cloud computing data auditing technologies mainly fall into two categories: cryptographic-based low-level possession proofs and artificial intelligence-based anomaly detection. However, facing increasingly sophisticated internal privilege abuse and advanced data poisoning attacks, existing technologies have the following shortcomings in practical applications:

[0003] 1. Traditional underlying cryptographic auditing has semantic blind spots, making it difficult to defend against legitimate insider poisoning attacks. Existing cloud auditing solutions mostly remain at the bit-level physical integrity verification. When an attacker uses a stolen privileged account to maliciously tamper with data through legitimate update commands or API requests (such as modifying the diagnosis results in an electronic medical record), the cloud storage system automatically generates a legitimate underlying hash signature for the updated data. This type of attack, where the physical data is not damaged but the business logic has been distorted, is difficult for traditional cryptographic auditing mechanisms to detect.

[0004] 2. Semantic detection based on pure AI models lacks cryptographic certainty, making it difficult to form a closed loop of audit evidence. Some existing studies have attempted to introduce models such as graph neural networks (GNNs) for semantic anomaly detection. However, such solutions have limitations: first, cloud data is massive and changes frequently, and the high computational cost of large models makes it difficult to apply to routine cloud audits; second, the inference results of AI models are essentially probability distributions, which cannot serve as mathematically deterministic cryptographic audit evidence when facing strict compliance reviews.

[0005] 3. Conventional sliding window-based dynamic threshold defense mechanisms are easily bypassed by slow, incremental poisoning. Existing semantic anomaly defense mechanisms typically rely on the moving average of historical states to set dynamic thresholds. Attackers can implement slow, incremental poisoning attacks, making only extremely minor logical modifications to the data each time. Over time, the system's dynamic thresholds are gradually assimilated by dirty data, eventually causing the system to misclassify abnormal data as normal business updates. Summary of the Invention

[0006] Therefore, it is necessary to provide a cloud data semantics and integrity dual-layer auditing method and system based on hypergraph and zero-knowledge proof to address the above-mentioned technical problems, which can effectively deal with data poisoning attacks and improve data security.

[0007] Firstly, this application provides a two-layer auditing method for cloud data semantics and integrity based on hypergraphs and zero-knowledge proofs. This method includes:

[0008] The plaintext database to be outsourced is encrypted as a whole and divided into several data blocks. The ciphertext block set is obtained, and the homomorphic authentication tag corresponding to each ciphertext block is generated. An MB-Tree is built based on the ciphertext blocks to generate the underlying physical root hash.

[0009] Based on the business logic of extracting plaintext database, a hypergraph is constructed with business entities as nodes and business records as hyperedges. An association matrix is ​​constructed according to the inclusion relationship between business entities and business records. A matrix commitment tree is constructed based on the association matrix, and the root hash of the association matrix is ​​generated.

[0010] The embedding feature matrix of global nodes in the hypergraph is extracted using a hypergraph neural network. Based on the embedding feature matrix, the semantic feature vector corresponding to each hyperedge is extracted. Based on the semantic feature vector, a semantic Merkle tree is constructed to generate a semantic root hash.

[0011] After receiving a random audit challenge, zero-knowledge blinding processing is performed on several ciphertext blocks sampled from the random audit challenge to generate a mask commitment and a blinded ciphertext aggregate block. Based on the blinded ciphertext aggregate block, mask commitment, homomorphic authentication label and MB-Tree, a cloud-based zero-knowledge aggregate proof is generated for data integrity auditing.

[0012] For the modified target hyperedge, extract the modified semantic feature vector, retrieve the original security semantic feature vector and the baseline semantic feature vector, obtain the relative semantic offset and absolute semantic offset, and compare the relative semantic offset and absolute semantic offset with the preset single dynamic threshold and absolute tolerance upper limit value for semantic integrity auditing.

[0013] In one embodiment, constructing an association matrix based on the inclusion relationship between business entities and business records, and constructing a matrix commitment tree based on the association matrix, includes:

[0014] Define an association matrix if and only if business entities Included in business records At that time, the correlation matrix The element at the specified position has a value of 1, otherwise it has a value of 0;

[0015] Obtain the hash value of each column vector in the association matrix and construct the matrix commitment tree.

[0016] In one embodiment, the embedding feature matrix of the global nodes of the hypergraph is extracted, and the semantic feature vector corresponding to each hyperedge is extracted based on the embedding feature matrix, including:

[0017] Obtain the node degree diagonal matrix and hyperedge degree diagonal matrix from the correlation matrix;

[0018] The initial attribute feature matrix of the node is obtained, and a forward propagation of a single-layer hypergraph neural network is performed to extract the embedding feature matrix of the global node; the hypergraph neural network uses an activation function implemented with fixed-point arithmetic for feature extraction.

[0019] Based on the correlation matrix, the embedding feature matrix, and the hyperedge degree diagonal matrix, obtain the semantic feature vector corresponding to each hyperedge.

[0020] In one embodiment, zero-knowledge blinding processing is performed on several ciphertext blocks sampled in a random audit challenge to generate a masked commitment and a blinded ciphertext aggregate block, including:

[0021] Calculate the aggregated tag of the homomorphic authentication tag;

[0022] Randomly select a blinding factor, calculate the mask commitment, and calculate the random challenge hash based on the mask commitment;

[0023] Based on the random blinding coefficient assigned to each ciphertext block for sampling, the sampled ciphertext blocks, the random challenge hash, and the blinding factor, a blinded ciphertext aggregate block is generated.

[0024] In one embodiment, the method further includes the following steps before performing semantic integrity auditing:

[0025] Obtain from the updated hypergraph the local submatrix centered on the target hyperedge, the first Merkel proof path of the submatrix in the matrix commitment tree, the modified semantic feature vector, and the second Merkel proof path corresponding to the modified semantic feature vector;

[0026] The authenticity of local submatrices is verified using the root hash of the correlation matrix and the first Merkle proof path.

[0027] After the verification is passed, forward inference verification is performed locally using the local network weight parameters, the initial attribute feature matrix of the node, and the fixed-point quantization rules to obtain the expected feature vector.

[0028] Forced verification is performed using the expected feature vector and the modified semantic feature vector, and the second Merkel proof path is used to check whether it matches the semantic root hash; if so, the semantic integrity audit is initiated.

[0029] In one embodiment, extracting the modified semantic feature vector, retrieving the original secure semantic feature vector and the baseline semantic feature vector, and obtaining the relative semantic offset and absolute semantic offset include:

[0030] Retrieve the semantic feature vector of the target hyperedge after the last legal update, and use it as the safe semantic feature vector. Obtain the relative semantic offset based on the Euclidean distance between the modified semantic feature vector and the safe semantic feature vector.

[0031] The semantic feature vector of the target hyperedge during the initialization phase is retrieved and used as the reference semantic feature vector. The absolute semantic offset is obtained based on the Euclidean distance between the modified semantic feature vector and the reference semantic feature vector.

[0032] In one embodiment, comparing the relative semantic offset and the absolute semantic offset with a preset single dynamic threshold and an absolute tolerance upper limit for semantic integrity auditing includes:

[0033] When the relative semantic offset is greater than the single dynamic threshold, it is determined that a single mutation-type advanced logic poisoning attack has occurred.

[0034] When the absolute semantic offset exceeds the absolute tolerance limit, it is determined that a chronic progressive tampering attack has occurred.

[0035] When the relative semantic offset is no greater than the single dynamic threshold and the absolute semantic offset is no greater than the absolute tolerance limit, the target superedge is determined to be a legitimate business update.

[0036] Secondly, this application also provides a two-tiered auditing system for cloud data semantics and integrity based on hypergraphs and zero-knowledge proofs. This system includes:

[0037] The data owner is responsible for encrypting the plaintext database to be outsourced, dividing it into several data blocks, obtaining a set of ciphertext blocks, generating homomorphic authentication tags for each ciphertext block, constructing an MB-Tree based on the ciphertext blocks, and generating the underlying physical root hash. Based on the business logic extracted from the plaintext database, a hypergraph is constructed with business entities as nodes and business records as hyperedges. An association matrix is ​​constructed based on the inclusion relationship between business entities and business records. A matrix commitment tree is constructed based on the association matrix, and the association matrix root hash is generated. The hypergraph neural network is used to extract the embedding feature matrix of the global nodes of the hypergraph. Based on the embedding feature matrix, the semantic feature vectors corresponding to each hyperedge are extracted. A semantic Merkle tree is constructed based on the semantic feature vectors, and the semantic root hash is generated.

[0038] The cloud service provider is used to perform zero-knowledge blinding processing on several ciphertext blocks sampled from the random audit challenge after receiving the random audit challenge, generate mask commitment and blinded ciphertext aggregate block, and generate cloud-based zero-knowledge aggregate proof based on the blinded ciphertext aggregate block, mask commitment, homomorphic authentication label and MB-Tree;

[0039] Third-party auditors are used to initiate random audit challenges; perform data integrity audits based on cloud-based zero-knowledge aggregation proofs; for modified target hyperedges, extract the modified semantic feature vector, retrieve the original security semantic feature vector and baseline semantic feature vector, obtain the relative semantic offset and absolute semantic offset, and compare the relative semantic offset and absolute semantic offset with preset single dynamic thresholds and absolute tolerance upper limits for semantic integrity auditing.

[0040] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the aforementioned cloud data semantics and integrity two-layer auditing method based on hypergraphs and zero-knowledge proofs.

[0041] Fourthly, this application also provides a computer-readable storage medium. This computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps in the aforementioned cloud data semantics and integrity two-layer auditing method based on hypergraphs and zero-knowledge proofs.

[0042] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps in the aforementioned cloud data semantics and integrity two-layer auditing method based on hypergraphs and zero-knowledge proofs.

[0043] The aforementioned cloud data semantic and integrity dual-layer auditing method and system based on hypergraphs and zero-knowledge proofs encrypts the entire plaintext database to be outsourced, divides it into several data blocks, obtains a set of ciphertext blocks, generates homomorphic authentication tags corresponding to each ciphertext block, constructs an MB-Tree based on the ciphertext blocks, and generates the underlying physical root hash. Based on the extracted business logic from the plaintext database, a hypergraph is constructed with business entities as nodes and business records as hyperedges. An association matrix is ​​constructed according to the inclusion relationship between business entities and business records. A matrix commitment tree is constructed based on the association matrix, generating the association matrix root hash. The hypergraph neural network is used to extract the embedding feature matrix of global nodes in the hypergraph. Based on the embedding feature matrix, the semantic feature vectors corresponding to each hyperedge are extracted. A semantic Merkle tree is constructed based on the semantic feature vectors, generating the semantic root hash. Upon receiving a random audit challenge, zero-knowledge blinding processing is performed on several ciphertext blocks sampled from the random audit challenge, generating a masked commitment and a blinded ciphertext aggregate block. Based on the blinded ciphertext aggregate block, masked commitment, homomorphic authentication label, and MB-Tree, a cloud-based zero-knowledge aggregate proof is generated for data integrity auditing. For the modified target hyperedge, the modified semantic feature vector is extracted, and the original security semantic feature vector and baseline semantic feature vector are retrieved to obtain relative and absolute semantic offsets. These relative and absolute semantic offsets are compared with preset single-time dynamic thresholds and absolute tolerance upper limits for semantic integrity auditing. This method proposes a two-layer integrity audit architecture that decouples physical holding from higher-order semantics, achieving responsibility isolation and precise tracing between low-level physical attacks and high-level business poisoning under legitimate identities. It also proposes a deterministic cryptographic verification closed loop based on fixed-point quantization and local forward recalculation; and a dual-threshold anti-chronic poisoning mechanism based on absolute baselines and single-time relative offsets is proposed to improve the stability of the cloud-based data semantic auditing process. Attached Figure Description

[0044] Figure 1 This is a system model diagram of a cloud data semantics and integrity dual-layer audit based on hypergraphs and zero-knowledge proofs in one embodiment. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0046] This application provides a two-layer auditing method for cloud data semantics and integrity based on hypergraphs and zero-knowledge proofs, such as... Figure 1 As shown, this method involves three types of core participating entities:

[0047] (1) Data Owner (DO): Represents the medical institutions within the region. Responsible for the local initialization of medical record data, including defining basic fields in the medical records, such as patient ID and diagnosis code, as business entities, and defining a single treatment session as a business record. Before data outsourcing, the DO is responsible for AES-256 encrypted segmentation of the medical records, constructing an initial topology based on a hypergraph, and generating absolute baseline features. The root hash is constructed using three tree structures: physical, semantic, and matrix. The ciphertext and credentials are then outsourced to the cloud, and the local plaintext is destroyed.

[0048] (2) Cloud Service Provider (CSP): Represents the cloud server storing medical record data. It is responsible for the physical storage of massive amounts of encrypted medical data and simultaneously maintains the dynamic Merkle B+ Tree (MB-Tree), Semantic Merkle Tree (SMT), and matrix commitment tree. Upon receiving medical record update instructions, such as when a doctor adjusts a prescription, the CSP must cooperate with the auditor to generate audit evidence containing proof of physical possession and a local business topology matrix.

[0049] (3) Third-Party Auditor (TPA): Represents an independent health regulatory auditing body. Commissioned by the DO, it regularly initiates audit challenges and is responsible for performing a two-tiered audit of data integrity and semantic anomalies. The TPA stores the absolute baseline features and root hash submitted by the DO, performs deterministic recalculation locally through fixed-point quantification rules, and, based on the dual threshold adjudication logic, blocks the subsequent reading and transfer of abnormal medical records in real time for mutation tampering and chronic poisoning.

[0050] The method includes the following steps:

[0051] S1. System initialization and dual-state commitment generation, this step includes:

[0052] S101, Ciphertext Generation, Cryptographic Parameters, and MB-Tree Initialization. Specifically, this includes:

[0053] (1) To ensure compatibility between fixed-length features and finite field operations, DO uses the secure symmetric encryption algorithm AES-256 to encrypt the entire plaintext database. Subsequently, the generated binary ciphertext string is divided into... A set of physical ciphertext blocks is obtained by dividing the data blocks into equal-length blocks. ,in , It is a large prime number.

[0054] (2) Key and parameter generation: DO selects bilinear mapping ,in For the order of The multiplicative cyclic group. For generators, DO selects random numbers as public parameters. As a private key Calculate the public key Define a cryptographic hash function. and standard collision-resistant hash functions .

[0055] (3) Homomorphic authentication tag generation: DO generates tags for each ciphertext block. Calculate its homomorphic authentication tag :

[0056]

[0057] in, This is a globally unique index for this data block.

[0058] (4) Constructing the MB-Tree: DO uses standard hash functions All As leaf nodes. A dynamic Merkle tree is constructed from bottom to top, where the value of a parent node is the secondary hash of the concatenation of the hash values ​​of all its child nodes. For example: Finally, the underlying physical root hash is calculated. and will Stored at TPA.

[0059] S102, Hypergraph Modeling of Business Topology and Semantic Merkle Tree (SMT) Initialization. Specifically, this includes:

[0060] (1) Hypergraph definition: Traditional graph structures can only express binary relationships between pairs of entities, while real cloud business records present higher-order multi-co-occurrence relationships. Therefore, this embodiment uses a hypergraph for modeling, with hyperedges completely defining all entities in the same business event.

[0061] Before data encryption was outsourced to the cloud, DO extracted the business logic from the plaintext database locally to build the initial hypergraph. Among them, nodes Represents all independent business entities globally. The total number of entities; superedge set Represents a complete business record that is independent of the overall system. To record the total number; It is a diagonal matrix representing the preset record confidence weights, which are initialized to the identity matrix I.

[0062] Define the hypergraph incidence matrix If and only if the entity Included in business records hour, Otherwise .

[0063] (2) Matrix Commitment: To prevent CSP from forging local topology structures later, DO extracts the association matrix column by column. Let the column vector corresponding to each hyperedge be denoted as , i.e., matrix The Columns. DO calculates the hash value for each column. Then, using these as leaf nodes, a matrix Merkle tree is constructed to generate the root hash of the associative matrix. And it was handed over to TPA for safekeeping.

[0064] (3) Degree matrix calculation: Calculate the degree diagonal matrix of the nodes. and hyperdiagonal matrix ,in:

[0065]

[0066] in, Accumulated entities The total number of records indicates the activity frequency of the entity; Accumulated records The total number of entities contained in the record represents the complexity of a single record.

[0067] (4) Hypergraph Feature Extraction Based on Fixed-Point Quantization: To ensure that TPA and CSP compute absolutely consistent semantic features under different hardware architectures, the system introduces fixed-point quantization rules. Let the initial attribute feature matrix of the node be... Perform forward propagation of a single-layer hypergraph neural network (HGNN) to extract the global node embedding feature matrix. ,in Feature dimensions output by the model:

[0068]

[0069] in, This is an activation function implemented using fixed-point arithmetic. These are the projection weights of network features with fixed parameters. The above formula uses left and right multiplication... and Perform double normalization to prevent high-frequency active entities from causing gradient explosion or masking the semantic features of long-tail entities during feature aggregation.

[0070] (5) Hyperedge semantic vector convergence: Extract the overall semantic features for a specific record (hyperedge). Calculate the semantic features of each hyperedge. semantic feature vector :

[0071]

[0072] in, For matrix The Row vectors. To prevent subsequent slow poisoning, DO will set... As an absolute benchmark feature, it is securely submitted to the TPA.

[0073] (6) Constructing the SMT semantic tree: DO hashes each feature vector As leaf nodes, semantic Merkle trees are constructed using the same bottom-up hash concatenation rules as MB-Tree, generating semantic root hashes. It is handed over to TPA for storage.

[0074] After completing the above initialization, DO will ciphertext set Tag set Correlation matrix The complete structures of MB-Tree, SMT, and matrix commitment trees are uploaded to the CSP. DO can securely destroy local plaintext data; the system's security relies on the keys held by the TPA and multiple state roots (…). We will jointly safeguard this.

[0075] S2, First-level audit: Data integrity audit based on ZK-DPDP (Zero-Knowledge Dynamic Provable DataPossession).

[0076] S201, Initiate a random audit challenge.

[0077] TPA generates a set of random audit challenges. .in Let the set of indexes for the sampled data blocks be denoted as . ; The random blinding coefficient is assigned to each sampled data block. TPA will Send to CSP.

[0078] S202, cloud-based zero-knowledge aggregation proof generation. After receiving the challenge, the CSP must prove that it has intactly preserved the ciphertext block. The following proof steps are to be implemented:

[0079] (1) Aggregated Tag: The aggregated tag for calculating homomorphic authentication tags. .

[0080] (2) Zero-knowledge blinding: CSP randomly selects blinding factors Calculate the mask commitment Calculate the hash of the random audit challenge. .

[0081] (3) Generate blinded ciphertext aggregation blocks:

[0082]

[0083] CSP will prove the evidence. Returned to TPA. Among them... It is the set of path hashes of the sampled nodes in the MB-Tree.

[0084] S203, TPA received supporting evidence Then, the bilinear pairing property is used to verify whether the following equation holds strictly:

[0085]

[0086] Formula derivation and proof:

[0087] The right side of the equation

[0088]

[0089]

[0090]

[0091] If the equation is not true or Unable to reconstruct If the underlying encrypted data is corrupted or lost, an abnormal alert will be issued directly; if the alert is valid, it proves that the cloud-based encrypted block is intact and will proceed to the second layer of auditing.

[0092] S3, Second-level audit: Semantic integrity audit based on absolute offset in high-dimensional space.

[0093] Assuming in At that moment, an internal privileged account initiated a legitimate UPDATE request, modifying the record. At this point, the first layer of auditing is passed, and the audit proceeds to the second layer, semantic integrity auditing. The specific steps of the second layer of auditing are as follows:

[0094] S301, Semantic Local Reconstruction and Evidence Submission.

[0095] CSP extracts the modified records based on the updated topology. Local correlation submatrix centered on CSP submits local submatrices to TPA. and its Merkel proof path in the matrix commitment tree (i.e., the first Merkel proof path), and the latest target hyperedge eigenvector. and its SMT verification path (i.e., the second Merkel proof path).

[0096] S302, Local Forward Reasoning and Authenticity Verification

[0097] After receiving the data, TPA first uses the matrix root hash. and path check The authenticity of the local matrix is ​​ensured to guarantee that it truly originates from the global topology recognized by the system.

[0098] After successful verification, TPA utilizes the locally stored network weights. Initial node features And strict fixed-point quantification rules, only applicable locally. Perform lightweight local forward inference verification to independently compute the expected feature vector. TPA mandatory verification and utilize Refactoring check for match .

[0099] S303, Calculation of bidirectional semantic offset distance.

[0100] After successful verification, TPA calculates the relative and absolute offsets to quantify the extent of logical changes caused by this update. The calculation process is as follows:

[0101] (1) Calculation of relative semantic offset:

[0102] TPA retrieves the security semantic feature vector of the business record stored in local storage after the last legitimate update, denoted as . Calculate the current update vector. and Euclidean distance in high-dimensional space, i.e., relative semantic offset :

[0103]

[0104] This metric is used to measure the magnitude of logical fluctuations caused by a single business operation.

[0105] (2) Calculation of absolute semantic offset:

[0106] TPA further retrieves the semantic feature vector of the business record stored during the system initialization phase, as the baseline semantic feature vector, denoted as... Calculate the current update vector. Compared with the baseline semantic feature vector The Euclidean distance, i.e., the absolute semantic offset.

[0107]

[0108] This metric measures the cumulative logical change in the record since the system was established.

[0109] S304, Dual Threshold Decision and Anti-Chronic Poisoning Judgment.

[0110] To defend against chronic poisoning attacks where attackers make only minor modifications each time, attempting to gradually assimilate the dynamic threshold over time, the system is configured with a single dynamic threshold. With absolute tolerance upper limit Dual adjudication mechanism:

[0111] (1) Dynamic single threshold calculation: ,in and This represents the mean and standard deviation of the set of relative distances from each valid single update in the history of this superedge. This threshold is dynamically adjusted based on normal business fluctuations and is used to constrain... .

[0112] (2) Absolute upper limit setting: This is a pre-defined absolute drift limit constant for the system's business logic, unaffected by subsequent data changes, used to constrain [data flow]. .

[0113] (3) Enforcement of the ruling rules:

[0114] like If a single mutation-type advanced logic poisoning attack occurs, the TPA will trigger a second-level red alert. The system will automatically freeze the subsequent reading and transfer of the contaminated data, effectively blocking the attack, and report the anomaly to the DO.

[0115] like If a chronic, progressive data corruption attack occurs, the TPA will trigger a second-level red alert. The system will automatically freeze the subsequent reading and transfer of the corrupted data, effectively blocking the attack, and will report the anomaly to the DO.

[0116] If and only if and When the system determines that the UPDATE operation does not destroy the hypergraph topology invariants between business entities, it is considered a legitimate business change.

[0117] Finally, after both layers of auditing passed, TPA updated its local security baseline status to [current state]. And update the hash tree root.

[0118] S4. System Status Dynamic Synchronization and Security Update Mechanism

[0119] After determining that the business change is legal, the system must perform strict two-level dynamic status synchronization updates to ensure the freshness and accuracy of subsequent audit evidence:

[0120] (1) Underlying physical state update: CSP uses the latest ciphertext block generated by the UPDATE command in the cloud to replace the old ciphertext block, and strictly follows the bottom-up hash concatenation rules to recalculate the node hash of the affected branch in MB-Tree until a brand new underlying physical root hash is generated. .

[0121] (2) High-level semantic state update: CSP updates the local association topology matrix according to the updated local association topology matrix. Synchronously update the affected column vector hashes in the association matrix commitment tree, and generate a new matrix root hash from the bottom up. Simultaneously, using updated and verified valid feature vectors... Update the leaf nodes and branches of the Semantic Merkle Tree (SMT) to generate a new semantic root hash. .

[0122] (3) Trusted Node State Persistence and Baseline Iteration: TPA receives the new physical root, new matrix root, and new semantic root submitted by CSP, along with their corresponding verification paths. After verifying the local verification tree structure is correct, TPA uses the new root hash set. The security coverage extends to the old state. Meanwhile, to ensure the continued operation of the anti-chronic poisoning mechanism, TPA will incorporate the latest validated feature vectors from the second-level audit. Assigning values ​​to the relative historical baseline state, i.e., execution. The absolute reference vector during system initialization The data remains tamper-proof. At this point, the system has completed a full and secure data dynamic flow and auditing loop.

[0123] The core innovations and beneficial effects of this invention are as follows:

[0124] 1. A two-layer integrity audit architecture that decouples physical possession from high-order semantics is proposed. This invention combines zero-knowledge dynamic data possession proof of the underlying ciphertext block with high-order multi-dimensional business logic topology modeling based on hypergraphs. This overcomes the vulnerability of traditional cryptographic auditing, which can only verify physical existence but cannot prevent tampering with legitimate permissions. It achieves responsibility isolation and accurate tracing between underlying physical attacks and high-level business poisoning under legitimate identities.

[0125] 2. A deterministic cryptographic verification closed loop based on fixed-point quantization and local forward recalculation is proposed. This invention abandons the traditional approach of using AI probabilistic outputs as black-box evidence, introduces an association matrix commitment mechanism, and transforms the feature extraction process of hypergraph neural networks into a fixed-point quantization forward recalculation with strict consistency. During auditing, the auditor only uses a local submatrix constrained by Merkle tree commitments to independently perform lightweight inference locally, avoiding the evidence chain break caused by cloud-based forged features, and transforming probabilistic semantic detection into an anti-counterfeiting verification closed loop based on hash commitments and deterministic computation.

[0126] 3. A dual-threshold anti-chronic poisoning mechanism based on absolute benchmark and single relative offset is proposed. This invention anchors the absolute initial feature vector of the business topology during the system initialization phase. By simultaneously constraining the single relative offset threshold based on the preceding legitimate state (preventing single mutation) and the cumulative drift distance upper limit based on the absolute benchmark (preventing long-arm assimilation) during auditing, it avoids attackers exploiting sliding window vulnerabilities to perform long-term, repeated, and minor legitimate logical tampering. This can, to a certain extent, reduce the risk of progressive semantic tampering being assimilated by dynamic thresholds and improve the stability of the cloud data semantic auditing process.

[0127] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0128] Based on the same inventive concept, this application also provides a cloud data semantic and integrity two-layer auditing system based on hypergraphs and zero-knowledge proofs. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the cloud data semantic and integrity two-layer auditing system based on hypergraphs and zero-knowledge proofs provided below can be found in the limitations of the cloud data semantic and integrity two-layer auditing method based on hypergraphs and zero-knowledge proofs described above, and will not be repeated here.

[0129] In one embodiment, such as Figure 1 As shown, a cloud data semantic and integrity two-layer auditing system based on hypergraph and zero-knowledge proof is provided, including: a data owner, used to encrypt the plaintext database to be outsourced and divide it into several data blocks, obtain a set of ciphertext blocks, generate homomorphic authentication tags corresponding to each ciphertext block, construct an MB-Tree based on the ciphertext blocks, and generate the underlying physical root hash; constructing a hypergraph with business entities as nodes and business records as hyperedges based on the extracted business logic of the plaintext database, constructing an association matrix according to the inclusion relationship between business entities and business records, constructing a matrix commitment tree based on the association matrix, and generating the association matrix root hash; using a hypergraph neural network to extract the embedding feature matrix of global nodes of the hypergraph, extracting the semantic feature vectors corresponding to each hyperedge based on the embedding feature matrix, constructing a semantic Merkle tree based on the semantic feature vectors, and generating the semantic root hash;

[0130] The cloud service provider is used to perform zero-knowledge blinding processing on several ciphertext blocks sampled from the random audit challenge after receiving the random audit challenge, generate mask commitment and blinded ciphertext aggregate block, and generate cloud-based zero-knowledge aggregate proof based on the blinded ciphertext aggregate block, mask commitment, homomorphic authentication label and MB-Tree;

[0131] Third-party auditors are used to initiate random audit challenges; perform data integrity audits based on cloud-based zero-knowledge aggregation proofs; for modified target hyperedges, extract the modified semantic feature vector, retrieve the original security semantic feature vector and baseline semantic feature vector, obtain the relative semantic offset and absolute semantic offset, and compare the relative semantic offset and absolute semantic offset with preset single dynamic thresholds and absolute tolerance upper limits for semantic integrity auditing.

[0132] The overall interaction process among the three is as follows:

[0133] Taking the medical field as an example, DO establishes a mapping relationship of "medical record fields - ciphertext blocks - hypergraph topology" to achieve cryptographic binding between business logic and physical storage, and completes the initial allocation of third-party keys and parameters. TPA periodically launches random audit challenges to CSP to perform the first layer of audit. CSP generates zero-knowledge proofs through homomorphic tag aggregation, and TPA verifies the objective physical existence of ciphertext blocks through bilinear pairing. When an internal privileged account initiates a legitimate UPDATE command to modify a medical record, the system must first ensure that the first layer of audit verification for the new ciphertext block passes. Subsequently, it automatically enters the second layer of semantic audit. CSP submits the updated local matrix and the double-verification path; TPA must first verify the authenticity of the matrix commitment, and then perform fixed-point forward inference recalculation to quantify the degree of logical change caused by this legitimate command. The system allows the medical record to flow only when the double threshold adjudication determines that the UPDATE operation is a normal business change. Meanwhile, CSP replaces the ciphertext at the underlying level and recalculates the affected tree branches; TPA securely overwrites the old state with the new root hash set and assigns the current feature to the relative historical benchmark, completing the closed loop of the audit evidence chain.

[0134] The modules in the aforementioned cloud data semantics and integrity dual-layer auditing system based on hypergraphs and zero-knowledge proofs can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can invoke and execute the corresponding operations of each module.

[0135] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in all of the above method embodiments.

[0136] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in all of the above method embodiments.

[0137] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in all of the above method embodiments.

[0138] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0139] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0140] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0141] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A two-layer auditing method for cloud data semantics and integrity based on hypergraphs and zero-knowledge proofs, characterized in that, The method includes: The plaintext database to be outsourced is encrypted as a whole and divided into several data blocks. A set of ciphertext blocks is obtained, and homomorphic authentication tags corresponding to each ciphertext block are generated. An MB-Tree is constructed based on the ciphertext blocks to generate the underlying physical root hash. Based on the business logic of extracting plaintext database, a hypergraph is constructed with business entities as nodes and business records as hyperedges. An association matrix is ​​constructed according to the inclusion relationship between the business entities and the business records. A matrix commitment tree is constructed according to the association matrix, and the root hash of the association matrix is ​​generated. The embedding feature matrix of the global nodes of the hypergraph is extracted using a hypergraph neural network. Based on the embedding feature matrix, the semantic feature vector corresponding to each hyperedge is extracted. Based on the semantic feature vector, a semantic Merkle tree is constructed to generate a semantic root hash. After receiving a random audit challenge, zero-knowledge blinding processing is performed on several ciphertext blocks sampled from the random audit challenge to generate a mask commitment and a blinded ciphertext aggregation block. Based on the blinded ciphertext aggregation block, the mask commitment, the homomorphic authentication tag, and the MB-Tree, a cloud-based zero-knowledge aggregation proof is generated for data integrity auditing. For the modified target hyperedge, extract the modified semantic feature vector, retrieve the original security semantic feature vector and the baseline semantic feature vector, obtain the relative semantic offset and the absolute semantic offset, and compare the relative semantic offset and the absolute semantic offset with the preset single dynamic threshold and the absolute tolerance upper limit value for semantic integrity auditing.

2. The method according to claim 1, characterized in that, The step of constructing an association matrix based on the inclusion relationship between the business entity and the business record, and constructing a matrix commitment tree based on the association matrix, includes: Define the association matrix if and only if the business entity Included in the business records At that time, the correlation matrix The element at the specified position has a value of 1, otherwise it has a value of 0; Obtain the hash value of each column vector in the association matrix and construct the matrix commitment tree.

3. The method according to claim 1, characterized in that, The step of extracting the embedding feature matrix of the global nodes of the hypergraph and extracting the semantic feature vector corresponding to each hyperedge based on the embedding feature matrix includes: Obtain the node degree diagonal matrix and the hyperedge degree diagonal matrix based on the correlation matrix; The initial attribute feature matrix of the node is obtained, and a forward propagation of a single-layer hypergraph neural network is performed to extract the embedding feature matrix of the global node; wherein the hypergraph neural network uses an activation function implemented with fixed-point arithmetic for feature extraction. Based on the correlation matrix, the embedding feature matrix, and the hyperedge degree diagonal matrix, the semantic feature vector corresponding to each hyperedge is obtained.

4. The method according to claim 1, characterized in that, The process of performing zero-knowledge blinding on several ciphertext blocks sampled for random audit challenges to generate masked commitments and blinded ciphertext aggregate blocks includes: Calculate the aggregated tag of the homomorphic authentication tag; Randomly select a blinding factor, calculate the mask commitment, and calculate a random challenge hash based on the mask commitment; The blinded ciphertext aggregate block is generated based on the random blinding coefficient assigned to each of the ciphertext blocks to be sampled, each of the ciphertext blocks to be sampled, the random challenge hash, and the blinding factor.

5. The method according to claim 1, characterized in that, Prior to performing semantic integrity auditing, the method further includes: Obtain from the updated hypergraph a local submatrix centered on the target hyperedge, the first Merkel proof path of the submatrix in the matrix commitment tree, the modified semantic feature vector, and the second Merkel proof path corresponding to the modified semantic feature vector; The authenticity of the local submatrix is ​​verified using the root hash of the correlation matrix and the first Merkel proof path; After the verification is passed, forward inference verification is performed locally using the local network weight parameters, the initial attribute feature matrix of the node, and the fixed-point quantization rules to obtain the expected feature vector. Forced verification is performed using the expected feature vector and the modified semantic feature vector, and the second Merkel proof path is used to check whether it matches the semantic root hash; if so, semantic integrity audit is initiated.

6. The method according to claim 1, characterized in that, The steps of extracting the modified semantic feature vector, retrieving the unmodified secure semantic feature vector and the baseline semantic feature vector, and obtaining the relative semantic offset and absolute semantic offset include: Retrieve the semantic feature vector of the target hyperedge after the last legal update, and use it as the safe semantic feature vector. Obtain the relative semantic offset based on the Euclidean distance between the modified semantic feature vector and the safe semantic feature vector. The semantic feature vector of the target hyperedge during the initialization phase is retrieved and used as the reference semantic feature vector. The absolute semantic offset is obtained based on the Euclidean distance between the modified semantic feature vector and the reference semantic feature vector.

7. The method according to claim 1, characterized in that, The step of comparing the relative semantic offset and the absolute semantic offset with a preset single dynamic threshold and an absolute tolerance upper limit for semantic integrity auditing includes: When the relative semantic offset is greater than the single dynamic threshold, it is determined that a single mutation-type advanced logic poisoning attack has occurred. When the absolute semantic offset is greater than the absolute tolerance upper limit, it is determined that a chronic progressive tampering attack has occurred. When the relative semantic offset is not greater than the single dynamic threshold and the absolute semantic offset is not greater than the absolute tolerance upper limit, the target hyperedge is determined to be a legitimate business update.

8. A cloud data semantic and integrity dual-layer auditing system based on hypergraphs and zero-knowledge proofs, characterized in that, The system includes: The data owner is responsible for encrypting the plaintext database to be outsourced, dividing it into several data blocks, obtaining a set of ciphertext blocks, generating homomorphic authentication tags corresponding to each ciphertext block, constructing an MB-Tree based on the ciphertext blocks, and generating the underlying physical root hash. Based on the business logic extracted from the plaintext database, a hypergraph is constructed with business entities as nodes and business records as hyperedges. An association matrix is ​​constructed according to the inclusion relationship between the business entities and the business records. A matrix commitment tree is constructed based on the association matrix, generating the association matrix root hash. An embedding feature matrix of the global nodes of the hypergraph is extracted using a hypergraph neural network. Semantic feature vectors corresponding to each hyperedge are extracted based on the embedding feature matrix. A semantic Merkle tree is constructed based on the semantic feature vectors, generating the semantic root hash. A cloud service provider, upon receiving a random audit challenge, performs zero-knowledge blinding processing on several ciphertext blocks sampled from the random audit challenge, generating a masked commitment and a blinded ciphertext aggregate block, and generates a cloud-based zero-knowledge aggregated proof based on the blinded ciphertext aggregate block, the masked commitment, the homomorphic authentication label, and the MB-Tree; A third-party auditor initiates random audit challenges; performs data integrity audits based on the cloud-based zero-knowledge aggregation proof; for the modified target hyperedge, extracts the modified semantic feature vector, retrieves the original security semantic feature vector and the baseline semantic feature vector, obtains the relative semantic offset and the absolute semantic offset, and compares the relative semantic offset and the absolute semantic offset with the preset single dynamic threshold and the absolute tolerance upper limit for semantic integrity auditing.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images