A method and electronic device for inferring a manipulation subject and snapshotting verifiable evidence for a public blockchain address cluster

Abstract

The application discloses a kind of open blockchain address cluster-oriented control subject inference and verifiable evidence snapshot method and electronic equipment.The method normalizes blockchain account book event and observation data into fund flow item, extracts address cluster interaction feature vector, removes macroscopic network noise by environmental disturbance subspace and orthogonal projection operator, obtains environmental disturbance-related feature vector.A convex quadratic programming model is constructed for candidate correlation edge, and a standardized input summary is registered before model solving.Then, the model is solved to obtain the optimal solution of witness variables, and the witness variable strength and KKT residual are calculated.A three-element gate determination is performed according to the witness variable strength, KKT residual and the consistency of the standardized input summary and the registered summary before solving.The candidate correlation edge that passes the determination is combined to output the control subject inference result, and a hash chain evidence snapshot is generated.The application provides technical support for third-party institutions to implement zero-trust independent verification.

Description

Technical Field

[0001] This invention relates to the field of blockchain electronic evidence management technology, and in particular to a method and electronic device for inferring the controlling entity and taking verifiable data snapshots for public blockchain address clusters. Background Technology

[0002] Addresses, accounts, or contracts in public blockchain systems often appear in pseudo-anonymous form. Compliance, regulatory, risk monitoring, and security evidence collection systems need to infer on-chain address clusters that may be controlled by the same source of manipulation without directly relying on off-chain identity information, and retain computational evidence that can be verified.

[0003] "Manipulator inference" refers to inferring, based on on-chain behavioral characteristics, that multiple seemingly independent addresses or address clusters actually belong to the same behind-the-scenes entity (manipulator), thereby clustering and merging them into a larger set. Furthermore, since the inference results are often used for judicial evidence analysis or anomaly transaction tracking, it is essential to address the pain point that "pure black-box inference cannot meet the verifiability requirements in judicial evidence scenarios." Therefore, it is necessary to introduce "verifiable evidence snapshots," which cryptographically solidify the standardized input summary, key model parameters, mathematical convergence accuracy, and inference results during the inference process, ensuring that the evidence content is tamper-proof and detectable, and supporting independent third-party verification.

[0004] Existing methods typically rely on heuristic rules, supervised learning models, or general graph clustering frameworks. However, heuristic rules are poorly adaptable to novel obfuscation strategies; pure black-box models such as supervised learning suffer from feature drift and insufficient interpretability; and general graph clustering results often lack a chain of evidence linked to the input data, model parameters, and numerical convergence accuracy, making it difficult to meet the auditability requirements in third-party verification and judicial evidence collection scenarios.

[0005] Therefore, a feasible technical solution is needed that can form a closed loop between mathematical inference, input consistency verification, and cryptographic evidence preservation. Summary of the Invention

[0006] The purpose of this invention is to provide a method and electronic device for inferring the controlling entity and taking verifiable data snapshots for public blockchain address clusters, thereby overcoming the aforementioned limitations of the prior art. The various technical effects of the preferred solutions among the many technical solutions provided by this invention are detailed below.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a method for inferring the controlling entity and taking verifiable data snapshots for public blockchain address clusters, comprising: Statistical values ​​of address clusters on a preset associated evidence dimension are extracted from blockchain ledger event and network propagation observation data, and the interaction feature vectors of the current batch of address clusters are spliced ​​together; an environmental disturbance subspace and an orthogonal projection operator are constructed based on historical stable observation data, and the interaction feature vectors are projected onto the orthogonal complement space of the environmental disturbance subspace to obtain the environmental disturbance-free associated feature vectors. The observation matrix and observation target vector of the candidate associated edge are obtained based on the associated feature vector. Witness variables and slack variables are set, and the convex quadratic programming model corresponding to the candidate associated edge is constructed based on the two-sided inequality constraint of capital conservation and the feature identifiability inequality constraint. Before solving the convex quadratic programming model, the standardized input data corresponding to the candidate associated edges are deterministically encoded and hashed to generate and register the standardized input summary before solving. Solve the convex quadratic programming model corresponding to each candidate association edge to obtain the optimal solution of the witness variable. Calculate the strength of the witness variable based on the components of each witness variable and their corresponding importance weights in the optimal solution. Calculate the KKT residual based on the solution results. Based on the strength of the witness variable, the KKT residual, and the consistency between the standardized input summary of the gate check and the standardized input summary before the solution, a ternary gate check is performed on the candidate associated edges. A unified association graph is constructed for all candidate association edges that pass the ternary gating judgment. Clustering and merging are performed on interconnected address clusters, the inference results of the manipulator are output, and a hash chain evidence snapshot is generated. The hash chain evidence snapshot is used by the verifier to regenerate the standardized input summary for verification, re-execute the solution of the convex quadratic programming model, and verify the newly generated standardized input summary for verification, witness variable strength, KKT residual, address cluster merging result, and hash chain integrity based on the candidate associated edges to be verified and the corresponding data in the evidence snapshot.

[0008] In some embodiments, the step of constructing an environmental disturbance subspace and an orthogonal projection operator based on historical stable observation data, and projecting the interactive feature vector onto the orthogonal complement space of the environmental disturbance subspace to obtain an environmentally disturbance-free associated feature vector includes: Singular value decomposition is performed on the matrix corresponding to the acquired historical stable observation data to obtain the left singular vector matrix, the singular value diagonal matrix, and the right singular vector matrix; The cutoff parameter r is determined based on the preset cumulative energy threshold, and the first r left singular vectors are selected to form a matrix. and the matrix The column space is used as the environmental perturbation subspace, and an orthogonal projection operator is constructed based on the environmental perturbation subspace; Subtract the historical stable observation mean vector from the interaction feature vector of the current batch address cluster, and then multiply it by the orthogonal projection operator to obtain the environmentally disturbed associated feature vector.

[0009] In some embodiments, constructing the convex quadratic programming model corresponding to the candidate associated edges includes: The objective function of the convex quadratic programming model includes an error penalty term calculated based on the observation matrix of the candidate associated edges and the observation target vector, a regularization smoothing term for the witness variables, and linear penalty terms corresponding to the capital conservation deviation slack variables and the feature identifiability slack variables. The constraints of the convex quadratic programming model include at least two-sided inequality constraints on capital conservation and identifiability constraints. The two-sided inequality constraints on capital conservation consist of a capital conservation deviation constraint matrix, a constraint objective vector, witness variables, a capital conservation error bound vector, and capital conservation deviation slack variables. The identifiability constraints on identifiability consist of a identifiability constraint matrix, a identifiability lower bound vector, witness variables, and identifiability slack variables.

[0010] In some embodiments, the witness variable includes multiple components, each of which corresponds to a multiple dimension of association evidence enabled by the current convex quadratic programming model. The dimensions of the associated evidence include the common interaction dimension, the change pattern dimension, the broadcast time sequence dimension, and the amount distribution dimension. The statistical value of the common interaction dimension is used to characterize the similarity between two address clusters in the same or adjacent batches when they jointly input, participate in transactions, aggregate, or split transactions. The statistical value of the change pattern dimension is used to characterize the similarity between two address clusters in terms of change address generation patterns, change ratios, or change address usage patterns. The statistical value of the broadcast time sequence dimension is used to characterize the similarity of the first observation time, arrival time difference, or propagation path of related transactions during network propagation. The statistical value of the amount distribution dimension is used to characterize the similarity between two address clusters in terms of inflow amount, outflow amount, net flow amount, or amount quantile distribution. Each component in the optimal solution of the witness variable is used to characterize the degree of evidence support for the inference that the address clusters at both ends of the candidate associated edge are controlled by the same manipulator, according to the corresponding associated evidence dimension.

[0011] In some embodiments, the generation and registration of the standardized input summary before solving includes: According to the preset deterministic coding rules, the standardized model input data and candidate associated edge basic data used to construct the convex quadratic programming model corresponding to the candidate associated edge are encoded and hashed to obtain the standardized input summary before the solution; the standardized input summary before the solution is recorded using the candidate associated edge identifier as an index.

[0012] In some embodiments, solving the convex quadratic programming model corresponding to each candidate associated edge, calculating the strength of the witness variable, and calculating the KKT residual based on the solution results include: Solve the convex quadratic programming model corresponding to the candidate association edge to obtain the optimal solution of the witness variable vector, wherein the components of the witness variable vector are arranged in a fixed order according to the association evidence dimension; Based on the optimal solution of the witness variable vector and the preset importance weight vector consistent with the fixed order of the associated evidence dimensions, the witness variable strength of the candidate associated edge is calculated to characterize the comprehensive degree of association between the address clusters at both ends of the candidate associated edge under the control of the same manipulating entity. The two-sided inequality constraint of capital conservation, the identifiability inequality constraint, and the variable range constraint are equivalently transformed into inequality constraints in a unified direction, and the KKT residual is calculated based on the solution results. The KKT residual is determined by the maximum residual among the original feasibility residual, the dual feasibility residual, the complementary relaxation residual, and the dual variable nonnegativity residual, and is used to characterize the numerical convergence accuracy of the convex quadratic programming model corresponding to the candidate associated edge.

[0013] In some embodiments, the determination condition corresponding to the ternary gating must simultaneously satisfy the following three conditions: The strength of the witness variable is greater than or equal to a preset association strength threshold; The KKT residual is less than or equal to a preset residual threshold; The gating verification normalized input summary generated from the current input is consistent with the normalized input summary before solving.

[0014] In some embodiments, outputting the inference result of the manipulator for the candidate associated edges determined by ternary gating and generating a hash chain evidence snapshot includes: Using the address clusters at both ends of the candidate association edges determined by the ternary gating as nodes, and the candidate association edges determined by the ternary gating as the connection relationship, an association graph is constructed. Clustering and merging are performed on the interconnected address clusters in the association graph to obtain merged address clusters. A merging result summary is generated based on the merged address clusters. The merged address clusters are used to represent the inference results of the manipulation subject. The candidate associated edge identifier, the address cluster identifiers at both ends of the candidate associated edge, the standardized input summary of the gated verification, the witness variable strength, the KKT residual, and the merged result summary are serialized according to a preset deterministic coding rule to obtain the snapshot payload. The standardized model input data used to construct the convex quadratic programming model corresponding to the candidate associated edge, the preset importance weight vector, the KKT residual calculation parameters, and the preset numerical tolerance used during the re-verification are serialized according to a preset deterministic coding rule to obtain the snapshot header metadata. The hash value of the preceding hash pointer, the hash value of the snapshot payload, and the hash value of the snapshot header metadata are concatenated and the hash is calculated again to generate the current hash pointer, which serves as the preceding hash pointer for the next evidence snapshot. The snapshot payload, the snapshot header metadata, and the hash pointer together constitute the hash chain evidence snapshot.

[0015] In some embodiments, the verification process also includes the following steps: The verification party regenerates the standardized input summary for verification based on the candidate associated edges to be verified and the corresponding data in the evidence snapshot, re-executes the solution of the convex quadratic programming model corresponding to the candidate associated edges, and performs consistency verification on the standardized input summary for verification, the strength of witness variables, KKT residuals and address cluster merging results. At the same time, it verifies the continuity and integrity of the hash chain to confirm that the input data, calculation process, numerical convergence status and evidence records corresponding to the inference results of the manipulating entity have not undergone unauthorized changes.

[0016] According to another aspect of the invention, an electronic device is also provided, comprising: at least one processor; at least one memory for storing processor-executable instructions; wherein the at least one processor implements, as described above, a method for inferring the controlling entity and taking verifiable data snapshots for a public blockchain address cluster by executing the executable instructions.

[0017] Implementing one of the above-described technical solutions of the present invention has the following advantages or beneficial effects: This invention isolates universal network noise from a mathematical and geometric perspective by introducing an environmental perturbation subspace and orthogonal projection, reducing erroneous associations caused by macro-level fluctuations in the blockchain. It abandons the black-box model and adopts a convex quadratic programming model with relaxation variables, transforming the manipulation entity's inference into a mathematical optimization problem with a globally optimal solution and KKT residual metric. This provides objective, transparent, and mathematically measurable technical evidence for auditing or judicial verification. By registering standardized input summaries before solving and implementing ternary gating based on witness variable strength, KKT residuals, and input summary consistency, it reduces the risk of incorrect merging due to input substitution or insufficient solution quality. By solidifying the gating verification of standardized input summaries, key model parameters, witness variable strength, and KKT residuals through hash chain evidence snapshots, it provides a traceable and verifiable computational evidence chain for third-party verification. The cross-auditing closed loop of "pre-summary + ternary gating + hash chain" not only reduces I / O congestion caused by invalid records but also ensures the integrity of the computational evidence chain and the irreversibility of the computation timeline, providing reliable technical support for third-party institutions to implement zero-trust independent verification. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a flowchart of a method for inferring the controlling entity and taking verifiable data snapshots for a public blockchain address cluster according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a system for inferring the controlling entity and replicable data snapshot for public blockchain address clusters, according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, various exemplary embodiments described below will be referenced to the accompanying drawings, which form part of the exemplary embodiments, illustrating various exemplary embodiments that may be used to implement the present invention. Unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. It should be understood that they are merely examples of processes, methods, and apparatuses consistent with some aspects of the present invention disclosed as detailed in the appended claims, and other embodiments may be used, or structural and functional modifications may be made to the embodiments listed herein without departing from the scope and spirit of the present invention.

[0020] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," etc., indicate the orientation or positional relationship based on the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the referred element must have a specific orientation, or be constructed and operated in a specific orientation. The terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. The term "multiple" means two or more. The terms "connected" and "linked" should be interpreted broadly, for example, they can be fixed connections, detachable connections, integral connections, mechanical connections, electrical connections, communication connections, direct connections, indirect connections through an intermediate medium, and can be the internal connection of two elements or the interaction relationship between two elements. The term "and / or" includes any and all combinations of one or more of the related listed items. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0021] To illustrate the technical solution described in this invention, specific embodiments are described below, showing only the parts related to the embodiments of this invention.

[0022] Example 1: like Figure 1 As shown in this embodiment, a method for inferring the controlling entity and taking verifiable data snapshots for public blockchain address clusters includes: S100. Extract statistical values ​​of address clusters on the preset associated evidence dimension from blockchain ledger event and network propagation observation data, and concatenate them to generate the interaction feature vector of the current batch of address clusters; construct an environmental disturbance subspace and orthogonal projection operator based on historical stable observation data, and project the interaction feature vector to the orthogonal complement space of the environmental disturbance subspace to obtain the environmental disturbance-free associated feature vector.

[0023] When executing S100, the blockchain ledger event and network propagation observation data are first normalized. Specifically, public blockchain ledger event and network propagation observation data are read in batches; a batch refers to a data processing unit divided according to a preset block height range, preset time window, or preset sampling period, and k represents the batch index. Ledger events can come from full nodes, archive nodes, verified offline ledger files, or compliant data interfaces. Addresses, accounts, or contracts are abstracted as nodes, and the set of addresses that has been merged according to deterministic rules is called an address cluster. The manipulation subject inference result output in this embodiment is limited to the inference result at the on-chain address cluster level and does not directly identify off-chain natural persons or institutions.

[0024] In some embodiments, blockchain ledger events are normalized into a unified data structure containing fund flow terms. For address cluster c in batch k, its normalized fund flow term is represented as: ; in, This represents the normalized cash flow term of address cluster c in batch k; This indicates the amount flowing in from outside address cluster c; This indicates the amount of money flowing out of address cluster c; This indicates the on-chain transaction fees or gas fees borne by address cluster c; This represents the block reward or protocol reward received by address cluster c; This indicates an adjustment item determined according to preset normalization rules. The adjustment item is used to handle batch ownership adjustments caused by block reorganization or transaction confirmation delays, or cross-chain bridge events or contract events that can be reliably mapped by preset normalization rules.

[0025] Cross-chain bridge events, off-chain transactions, or other non-standard events that cannot be reliably mapped to cash flow items by the pre-defined normalization rules are not included in the aforementioned normalized cash flow items. Instead, they are recorded as local cash conservation error boundaries according to the pre-defined error boundary generation rules. The local capital conservation error bound is used as a priori parameter to construct the capital conservation error bound vector corresponding to the candidate associated edges in subsequent steps. .

[0026] It is easy to understand that heterogeneous blockchain underlying data (such as UTXOs, account balances, and smart contract events) is extremely complex and cannot be directly calculated. This embodiment standardizes it by using unified fund flow terms and error bounds, providing a consistent and high-quality quantitative data source for subsequent "manipulation entity inference," thus laying the foundation for the entire inference process.

[0027] Furthermore, statistical values ​​corresponding to the dimensions of associated evidence are extracted from ledger events and network propagation observation data in a preset fixed order, and then concatenated to generate the interaction feature vector of the current batch of address clusters. The dimensions of associated evidence include the common interaction dimension, the change pattern dimension, the broadcast time sequence dimension, and the amount distribution dimension.

[0028] Among them, the statistical values ​​of the common interaction dimension are used to characterize the similarity of two address clusters in the same or adjacent batches in terms of common input, common participation in transactions, common aggregation, or common splitting; the statistical values ​​of the change pattern dimension are used to characterize the similarity of two address clusters in terms of change address generation rules, change ratio, or change address usage patterns; the statistical values ​​of the broadcast time sequence dimension are used to characterize the similarity of the first observation time, arrival time difference, or propagation path of related transactions during network propagation; and the statistical values ​​of the amount distribution dimension are used to characterize the similarity of two address clusters in terms of transfer-in amount, transfer-out amount, net flow amount, or amount quantile distribution. After normalization, the statistical values ​​corresponding to the above-mentioned correlation evidence dimensions are arranged or concatenated in a preset fixed order of common interaction dimension, change pattern dimension, broadcast time sequence dimension, and amount distribution dimension to form a column vector, which is the interaction feature vector of the address clusters in the current batch. .

[0029] The aforementioned similarity can be determined using existing technologies, such as cosine similarity or Euclidean distance transformation similarity, with the specific similarity calculation method determined according to preset feature calculation rules. The aforementioned normalization process can employ existing data standardization methods such as logarithmic scaling followed by normalization or max-min normalization to convert statistical values ​​of different dimensions into comparable numerical forms. The aforementioned similarity calculation method, normalization method, and preset fixed order remain consistent throughout the candidate association edge generation, convex quadratic programming model construction, model solving, ternary gating determination, and verification processes. This preset fixed order ensures consistency in the dimensional correspondences of the interaction feature vector, observation matrix, witness variable components, and evidence snapshot.

[0030] This interactive feature vector This still contains macroscopic network noise, which will serve as the basic input for the next step of extracting the environmentally disturbed associated feature vector.

[0031] To avoid misinterpreting macroscopic network noise such as network congestion, node geographic distribution, and sampling delay as evidence of address cluster association, this embodiment uses historical stable observation data to construct an environmental perturbation subspace. Specifically, Singular value decomposition is performed on the matrix corresponding to the acquired historical stable observation data to obtain the left singular vector matrix, the singular value diagonal matrix, and the right singular vector matrix. The cutoff parameter r is determined based on the preset cumulative energy threshold, and the first r left singular vectors are selected to form a matrix. and the matrix The column space is used as the environmental perturbation subspace, and an orthogonal projection operator is constructed based on the environmental perturbation subspace; Subtract the historical stable observation mean vector from the interaction feature vector of the current batch address cluster, and then multiply it by the orthogonal projection operator on the left to obtain the environmentally disturbed associated feature vector.

[0032] In a specific embodiment, the mathematical expression corresponding to the environmentally undisturbed associated feature vector obtained above is: , , ; Among them, a historical stable observation matrix is ​​constructed based on historical stable observation data. ,in, The number of dimensions for node interaction features. The number of historically stable observation batches or samples; singular value decomposition is performed on the historically stable observation matrix to obtain... ,in, It is a left singular vector matrix. It is a singular value diagonal matrix. It is a right singular vector matrix.

[0033] Sort the singular values ​​in descending order, and let the i-th... Each singular value is According to the preset cumulative energy threshold (Value) Determine the cutoff parameters (Select the smallest integer that satisfies the condition) such that ,in, This represents the total number of singular values. Before selection... A matrix is ​​formed by left singular vectors. and will The column space is used as the environmental perturbation subspace.

[0034] Construct orthogonal projection operators based on environmental perturbation subspace ,in, for An identity matrix. The interaction feature vectors of the current batch address clusters. Subtract the historical stable observation mean vector Then, the vectors are projected onto the orthogonal complement of the environmental perturbation subspace to obtain the environmentally perturbation-free associated feature vectors. .

[0035] To be easily understood, the purpose of calculating the above projection formula (i.e., orthogonal projection to remove environmental disturbances) is to define a universal macroscopic noise space across the entire network through singular value decomposition, and then use orthogonal projection to remove this noise. The calculated correlation feature vector after removing environmental disturbances... It will directly serve as the core data source for constructing the observation matrix of the convex quadratic programming model in subsequent steps, thereby reducing the probability of false associations caused by environmental disturbances from a mathematical and physical perspective.

[0036] S200. Based on the associated feature vectors, obtain the observation matrix and observation target vector of the candidate associated edges, set witness variables and slack variables, and construct the convex quadratic programming model corresponding to the candidate associated edges based on the two-sided inequality constraint of capital conservation and the feature identifiability inequality constraint.

[0037] Before constructing the convex quadratic programming model, candidate related edges are predefined, denoted as . ,in and These are two on-chain address clusters whose merging is pending verification. In practical terms, they represent a "highly probable but uncertain control relationship," namely, address clusters... and Superficially, they appear to be manipulated by the same behind-the-scenes entity. Candidate related edges constitute the "basic case-filing unit" for subsequent mathematical inference, which can be generated from statistical clues corresponding to one or more related evidence dimensions among the common interaction dimension, change-making pattern dimension, broadcast timing dimension, and amount distribution dimension. Only through solving the convex quadratic programming model and subsequent ternary gating will this "candidate" state be determined as a mergeable related state.

[0038] For each candidate associated edge Construct witness variable vectors according to a preset fixed order of related evidence dimensions. Witness variable vector For the variables to be solved in the convex quadratic programming model, each component corresponds to one of the following dimensions: common interaction dimension, change pattern dimension, broadcast time sequence dimension, and amount distribution dimension. These dimensions are used to characterize the degree of evidence support for the inference that the address clusters at both ends of the candidate related edge are controlled by the same controlling entity.

[0039] In one implementation, candidate associated edges The corresponding witness variable vector is represented as follows ,in, The witness component corresponding to the common interaction dimension. For the witness component corresponding to the change pattern dimension, The witness component corresponding to the broadcast time sequence dimension. This represents the witness component corresponding to the monetary distribution dimension. The observation matrix of the candidate association edge, the feature identifiability constraint matrix, and the preset importance weight vector used to calculate the strength of the witness variable are all arranged in the same dimensional order as described above.

[0040] In some embodiments, the objective function of the convex quadratic programming model includes an error penalty term calculated based on the observation matrix of the candidate associated edges and the observation target vector, a regularization smoothing term for the witness variables, and linear penalty terms corresponding to the capital conservation deviation slack variables and the feature identifiability slack variables.

[0041] The constraints of a convex quadratic programming model include at least two-sided inequality constraints related to capital conservation and identifiability constraints. The two-sided inequality constraints related to capital conservation consist of a capital conservation deviation constraint matrix, a constraint objective vector, witness variables, a capital conservation error bound vector, and capital conservation deviation slack variables. The identifiability constraints related to identifiability consist of a identifiability constraint matrix, a identifiability lower bound vector, witness variables, and identifiability slack variables. Specifically, For each candidate associated edge Construct the following convex quadratic programming model: Objective function: ; Capital conservation inequality constraints on both sides: ; Feature identifiability inequality constraints and variable range: ; in, This is the observation error penalty term, representing the deviation between the model prediction and the observed target; Regularized smoothing terms for witness variables to prevent witness variables from becoming too large or overfitting; For the linear penalty term of the slack variable of the conservation of funds deviation, The linear penalty term for the feature identifiability slack variable is used to prevent the model from easily violating the capital conservation constraint and the feature identifiability constraint.

[0042] Furthermore, the solution variables for the convex quadratic programming model are , , .

[0043] The witness variable vector represents the degree of evidence supporting the inference that the address clusters at both ends of a candidate association edge are controlled by the same entity, based on multiple dimensions of related evidence. Each component of the witness variable vector corresponds to one of the following dimensions: common interaction dimension, change pattern dimension, broadcast timing dimension, and amount distribution dimension; constraints. The evidence support value corresponding to each related evidence dimension is between 0 and 1.

[0044] For the slack variable of the money conservation deviation, it is used to allow the two-sided inequality constraint of money conservation to be relaxed to a limited extent outside the original money conservation error boundary in a penalized manner; These are feature identifiability relaxation variables, used to allow for limited relaxation of feature identifiability inequality constraints when outlier data is missing or observations are insufficient. , and They represent lengths of , and A column vector whose elements are all 1s.

[0045] Candidate edges The observation matrix, whose internal elements are composed of the aforementioned environmentally undisturbed correlated eigenvectors. It is constructed by mapping; in one implementation, for candidate associated edges Based on the preset feature index, the environmentally undisturbed associated feature vectors corresponding to each batch or time window are used. Extracting address clusters and The relevant feature observations are normalized when necessary; the feature observations are then arranged in a preset fixed order to form candidate related edges. observation matrix .

[0046] This is the observation target vector corresponding to the candidate association edge, used to represent the target association strength of the candidate association edge in the observation data. In one implementation, for the observation matrix... For each feature observation record, according to the preset target generation rules, a target association strength value is generated based on the observation values ​​of the common interaction dimension, change pattern dimension, broadcast time sequence dimension, and amount distribution dimension in that feature observation record. The target association strength values ​​are then arranged in the order of the observation records to obtain the observation target vector. The default target generation rules do not use the witness variables to be solved and remain consistent throughout the same solution and verification process.

[0047] Capital conservation deviation constraint matrix Constraint target vector and the error boundary vector of capital conservation The generation method is as follows: for candidate associated edges Constraint records are generated on a batch or time window basis, based on the capital conservation relationship. Each constraint record corresponds to... a line of An element and One of the elements.

[0048] For each capital conservation relationship, based on the batch or time window and the address cluster... and address cluster The relevant standardized cash flow items are used to generate constraint coefficient rows in a preset fixed order; each column in this constraint coefficient row corresponds to a witness variable vector. One of the witness components, the coefficient of which is unrelated to the conservation of funds relationship can be set to zero. The constraint matrix of the conservation of funds deviation is formed by arranging the constraint coefficient rows in order. .

[0049] Constraint target vector Each element in the equation is calculated from the normalized cash flow items involved in the same cash conservation relationship according to the preset cash conservation target generation rules, and is used to represent the observation target that the model prediction value should approach under the cash conservation relationship.

[0050] Capital conservation error boundary vector It is obtained by aggregating local capital conservation error bounds. Specifically, for each capital conservation relationship, the local capital conservation error bounds of the address clusters involved in the relationship within the corresponding batch or time window are read, and combined with sampling error, rounding error, or unnormalized event error, to generate... The corresponding non-negative elements. The local capital conservation error boundary is recorded in S100. Capital conservation error boundary vector Used to limit the allowable deviation range of the two-sided inequality constraint for the conservation of funds.

[0051] Feature Identifiability Constraint Matrix and feature discriminability lower bound vector The generation method is as follows: generated according to a preset fixed order of preset feature index and associated evidence dimension. Each row in the table represents a minimum identifiability requirement. If an identifiability requirement corresponds to a single dimension of related evidence, the row has a preset coefficient set in the corresponding dimension column and zero in the other dimension columns; if an identifiability requirement corresponds to a combination of multiple dimensions of related evidence, the row has a preset combination weight set in the corresponding multiple dimension columns. Each element in is The lower bound of the lowest distinguishability strength for the corresponding row. Feature distinguishability inequality constraint. This is used to require candidate related edges to reach identifiable strength in at least one or more preset dimensions of related evidence. Used to provide limited relaxation when anomalous data is missing or insufficient observations.

[0052] For the regularization parameters of the witness variable, and These are the penalty parameters for the slack variables related to the conservation of funds bias and the slack variables related to feature identifiability, respectively. >0, >0, >0. To control the smoothness of the witness variable, its value range is typically [value range missing]. Taking values ​​that are too large can lead to overly conservative inferences. and Tolerances for capital conservation and feature identification errors, as a relaxation penalty for hard constraints, typically need to be set much larger than... (For example This ensures that constraints are only compromised in extreme cases where data is missing.

[0053] To ensure dimension matching of all matrices and vectors in the above convex quadratic programming model, let... Candidate related edges The number of currently enabled dimensions of related evidence. The number of characteristic observation records used to construct the observation matrix. This represents the number of constraints corresponding to the law of conservation of funds. This represents the number of feature identifiability constraints. The witness variable vector satisfies... The slack variable satisfies the conservation of funds bias. The slack variables of feature identifiability satisfy The observation matrix and the observation target vector satisfy... , The constraint matrix for the capital conservation deviation, the constraint objective vector, and the capital conservation error boundary vector satisfy the following conditions: , , The feature discriminability constraint matrix and the feature discriminability lower bound vector satisfy... , .in, Represents the set of real numbers. Let represent the set of nonnegative real numbers.

[0054] The aforementioned observation matrix, observation target vector, capital conservation deviation constraint matrix, constraint target vector, capital conservation error bound vector, feature identifiability constraint matrix, feature identifiability lower bound vector, as well as model parameters and constraint parameters, all belong to the standardized model input data of the convex quadratic programming model corresponding to the candidate associated edges; the object of the convex quadratic programming model solution is the witness variable vector, capital conservation deviation slack variable, and feature identifiability slack variable, rather than the aforementioned matrix or vector.

[0055] To put it simply, the purpose of designing the aforementioned convex quadratic programming model and its constraint formulas is to transform the fuzzy correlation inference problem into a verifiable mathematical optimization problem. The generated objective vector and constraint matrix not only define the inference logic in the current step, but also enable the solution results to be repeatedly verified by a third party under the same input data, model parameters, and numerical tolerance conditions, thus providing a verifiable mathematical basis for subsequent gating.

[0056] This step abandons the uninterpretable black box of deep learning and constructs a convex quadratic programming model with slack variables for inference. Under the condition of fixed parameters, this model can obtain the global optimum, and the numerical convergence quality can be objectively measured by KKT residuals, thus laying the technical foundation for subsequent replicable data snapshots.

[0057] S300. Before solving the convex quadratic programming model, perform deterministic encoding and hash calculation on the standardized input data corresponding to the candidate associated edges, and generate and register the standardized input summary before solving.

[0058] In some embodiments, candidate associated edges The corresponding standardized input data before solving is denoted as The standardized input data before solving the candidate associated edges refers to the input data obtained from solving the candidate associated edges. Before solving the corresponding convex quadratic programming model, standardized model input data and candidate associated edge basic data are used to uniquely determine the convex quadratic programming model. The standardized input data includes at least the candidate associated edge identifier, the address cluster identifiers at both ends of the candidate associated edge, and the standardized model input data of the convex quadratic programming model described in S200.

[0059] Standardized input data before solving Serialization is performed according to a preset deterministic encoding rule, and a cryptographic hash value is calculated to obtain a standardized input digest before solving: ; in, Candidate related edges Standardized input summary before solving; The default cryptographic hash function; A deterministic encoding function for standardized input data.

[0060] Deterministic encoding rules should include at least the following: encoding according to a preset field list and a fixed field order. Each field in the input is encoded; for numeric fields, fixed numerical precision, fixed rounding method, and fixed byte order are used; for set fields, they are sorted according to a preset sorting key before encoding; for matrix fields, the matrix identifier, number of matrix rows, number of matrix columns, and matrix elements arranged in a preset row and column order are encoded simultaneously; for vector fields, the vector identifier, vector length, and vector elements arranged in a preset element order are encoded simultaneously. Thus, the same standardized input data can generate a consistent standardized input summary at different execution devices or different verification entities. It should be noted that the deterministic encoding rules in subsequent steps are consistent with those in this step.

[0061] During the registration process before solving, the candidate associated edge is recorded using its identifier as an index field. The corresponding standardized input summary before solving The registered information can be written to persistent storage with access control.

[0062] This step packages the standardized input data into a hash digest before solving the convex quadratic programming model, providing a tamper-detectable "pre-data fingerprint" for the "reproducible data snapshot".

[0063] S400. Solve the convex quadratic programming model corresponding to each candidate associated edge to obtain the optimal solution of the witness variable. Calculate the strength of the witness variable based on the components of each witness variable and their corresponding importance weights in the optimal solution, and calculate the KKT residual based on the solution results.

[0064] In some embodiments, for each convex quadratic programming model corresponding to a candidate associated edge constructed in step S200, a preset solver is used to perform the solution, obtaining the optimal solutions for the witness variable, the capital conservation deviation slack variable, and the feature identifiability slack variable. The preset solver can be implemented using the interior-point method, the active set method, the operator splitting method, or other existing numerical optimization algorithms suitable for convex quadratic programming problems. When the convex quadratic programming models corresponding to multiple candidate associated edges are independent or approximately independent, parallel or batch solving methods can be used. After the solution is completed, the KKT residuals, used to measure the accuracy of numerical convergence, are calculated.

[0065] When solving, the convex quadratic programming model is uniformly represented in a standard form: ; in, , indicating that the witness variable , Capital conservation bias slack variables and feature identifiability slack variables A unified decision variable vector formed by stacking columns; The objective function of the convex quadratic programming model in step S200; To unify the inequality constraint matrix, To unify the constraints on the right-hand side of the inequality vector.

[0066] Unified Inequality Constraint Matrix and the right-hand side vector of the unified inequality constraint The constraints are obtained by equivalent transformation and row-by-row superposition of the two-sided inequality constraints related to capital conservation, the identifiability inequality constraints, and the variable range constraints from step S200. Specifically, for inequality constraints in the form of "greater than or equal to", they are transformed into "less than or equal to" form by multiplying both sides by negative numbers. For nonnegative variable constraints, they are also equivalently transformed into "less than or equal to" form. For two-sided inequality constraints, they are split into upper bound constraints that limit the expression to no more than the upper limit value, and lower bound constraints that limit the expression to no less than the lower limit value. The upper bound constraints are directly incorporated into the inequality constraints in the "less than or equal to" direction after rearranging terms, and the lower bound constraints are equivalently transformed into inequality constraints in the "less than or equal to" direction after multiplying both sides by negative numbers, and then superimposed with the upper bound constraints row-by-row. The equivalent transformation does not change the feasible region of the convex quadratic programming model in step S200; it is only used to unify various constraints into a standard form suitable for solving and calculating KKT residuals.

[0067] and These are the equality constraint matrix and vector, respectively. Based on the construction in step S200, the capital conservation relationship has been represented by two-sided inequality constraints and slack variables. This embodiment does not separately set equality constraints, therefore... and It can be viewed as an empty matrix and an empty vector, corresponding to the dual variables of the equality constraint. Pick .

[0068] In this solution process, the optimal solution for the unified decision variables is obtained. and from unified decision variables The witness variable vector corresponding to the candidate associated edge is obtained from the optimal solution. The optimal solution. The witness variable vector includes multiple components, which are arranged in a fixed order according to the common interaction dimension, change pattern dimension, broadcast time sequence dimension, and amount distribution dimension, and represent the evidence support value of the corresponding related evidence dimension.

[0069] In obtaining the witness variable vector After finding the optimal solution, candidate associated edges are calculated according to the following formula. Witness variable strength: ; in, Candidate related edges The witness variable strength is used to characterize the comprehensive degree of association between the address clusters at both ends of the candidate association edge, which are controlled by the same manipulating entity; j is a preset importance weight vector consistent with the fixed order of the association evidence dimensions, and its components correspond to the common interaction dimension, change pattern dimension, broadcast time sequence dimension, and amount distribution dimension, respectively, and each component is non-negative and satisfies This is used to characterize the relative importance of different dimensions of related evidence in the calculation of overall strength.

[0070] In addition, KKT residuals are calculated simultaneously based on the solution results. The KKT residuals are determined by the maximum value among the original feasibility residuals, dual feasibility residuals, complementary relaxation residuals, and dual variable nonnegativity residuals, and are used to characterize the numerical convergence accuracy of the convex quadratic programming model corresponding to the candidate associated edges.

[0071] Calculate the following KKT residuals: , , , , ; in, , , and These are the original feasibility residual, the dual feasibility residual, the complementary relaxation residual, and the non-negativity residual of the dual variable, respectively. The objective function of the convex quadratic programming model exist The gradient vector at that point; For the dual variables of the inequality constraints, These are the dual variables of the equality constraints, and both are output by the preset solver; This indicates taking the non-negative part of each element; Represents element-wise product; This represents the infinite norm. The parameters used to calculate the KKT residuals are collectively referred to as KKT residual calculation parameters.

[0072] This step is the execution engine of the system's computing power. It obtains the optimal solution of the witness variable vector corresponding to the candidate associated edge based on the solution result of the convex quadratic programming model, and further calculates the witness variable strength and KKT residual; among them, the witness variable strength is used to measure the comprehensive association degree of the candidate associated edge, and the KKT residual is used to measure the numerical convergence accuracy of the model solution. Both serve as the input basis for the subsequent ternary gating decision.

[0073] S500. Based on the strength of the witness variable, the KKT residual, and the consistency between the standardized input summary and the standardized input summary before the solution, perform ternary gating on the candidate associated edges.

[0074] In some embodiments, the decision condition for performing the ternary gating decision must simultaneously satisfy the following three conditions: The witness variable strength is greater than or equal to the preset association strength threshold; KKT residuals are less than or equal to a preset residual threshold; The gating verification normalized input summary generated from the current input is consistent with the normalized input summary before solving.

[0075] If any one of the conditions is not met, the candidate associated edge will not enter the submission state. Only when all three conditions are met simultaneously will the candidate associated edge enter the submission state and be used for subsequent generation of a complete evidence snapshot and merging of the corresponding address clusters.

[0076] Three-element gating result Expressed as a formula: ; in, For the correlation strength threshold, The range of values ​​is usually set to This is used to filter out candidate address pairs with high correlation strength and prevent coincidental associations caused by weak features; The residual threshold, The numerical convergence accuracy is determined, and the range of values ​​is usually within a certain range. between; A cryptographic hash function consistent with S300; Standardized input data for performing gating verification (standardized input processing is consistent with S300); For the deterministic coding function in S300; Candidate related edges in S300 A standardized input summary before solving. Only. Only then will it enter the submission state and generate a complete evidence snapshot.

[0077] The purpose of calculating the strength of witness variables and the ternary gating formula is to construct a rigorous evidence admission verification mechanism. Only candidate association edges that simultaneously meet the following requirements can pass: strong substantive relevance (witness variable strength meets the standard), high calculation accuracy (residual compliance), and unaltered input data (digest consistency). This filters out weak associations and invalid persistence overhead, and ensures highly reliable evidence records for subsequent snapshot evidence storage.

[0078] This step is a rigorous check on the transformation of inference results into evidence. It constructs a ternary gating system using the strength of witness variables (how strong is the substantial correlation), KKT residuals (how accurate are the calculations), and input summary consistency (whether there has been any substitution). It ensures that only candidate correlation edges that simultaneously satisfy substantial correlation, numerical convergence, and input consistency enter the subsequent evidence snapshot solidification process.

[0079] S600: Construct a unified association graph for all candidate association edges that have passed the ternary gating judgment, perform clustering and merging on interconnected address clusters, output the inference results of the manipulator, and generate a hash chain evidence snapshot.

[0080] For candidate association edges that pass the ternary gating test, the address clusters at both ends of the candidate association edge are determined as mergeable address clusters. An association graph is constructed using the address clusters at both ends of the candidate association edges that pass the ternary gating test as nodes and the candidate association edges that pass the ternary gating test as connections. Clustering and merging are performed on the interconnected address clusters. If multiple address clusters are interconnected through one or more candidate association edges that pass the ternary gating test, these multiple address clusters are merged into a single merged address cluster. The merged address cluster is used to represent the inference results of the manipulating entity. For each merged address cluster, its original address cluster identifiers and member addresses are deduplicated, sorted, and serialized and hashed using a preset deterministic encoding rule to generate a merged result digest. The merged result digest is used by a third party to verify whether the address cluster merging results are consistent.

[0081] In some embodiments, generating a hash chain evidence snapshot includes at least: After deterministically sorting each candidate association edge determined by the ternary gating according to its identifier, corresponding evidence snapshots are generated sequentially and combined using a hash chain. The specific steps are as follows: The candidate associated edge identifier, the address cluster identifiers at both ends of the candidate associated edge, the gated verification standardized input digest, the witness variable strength, the KKT residual, and the merged result digest are serialized using a preset deterministic coding rule to obtain the snapshot payload. And store; The standardized model input data used to construct the convex quadratic programming model corresponding to the candidate associated edges, the preset importance weight vector, the KKT residual calculation parameters, and the preset numerical tolerance used in the verification are serialized using preset deterministic encoding rules to obtain the snapshot header metadata. And store; The hash value is calculated by concatenating the previous hash pointer, the hash value of the snapshot payload, and the hash value of the snapshot header metadata, and the current hash pointer is generated and fixed. The current hash pointer serves as the previous hash pointer for the next evidence snapshot. The previous hash pointer for the first evidence snapshot is a preset initial hash value.

[0082] ; in, For snapshot sequence number; For the first One hash pointer; This is a pointer to the previous hash. This indicates byte string concatenation.

[0083] The purpose of calculating this hash chain formula is to leverage the cryptographic avalanche effect to lock the current computational evidence against the preceding context. The calculated hash pointers... This will serve as an anti-counterfeiting label inferred by the entity responsible for the manipulation, allowing for rigorous non-repudiation verification during subsequent third-party audits. If any snapshot payload or parameters are modified, subsequent hash chain verification will fail.

[0084] The snapshot payload and snapshot header metadata are used to form the content body of the evidence snapshot. The hash pointer is bound to the content body of the evidence snapshot, and together they form a hash chain evidence snapshot and are stored.

[0085] Establish the association between the merged address cluster, the merge result summary, and the candidate associated edges that triggered the merge of the address cluster and their evidence snapshot set.

[0086] This step directly addresses the "evidence snapshot" topic. It solidifies the relevant underlying evidence data in the form of a context-bound hash chain. This cryptographic structure ensures that any minor tampering with the historical inference process will cause subsequent verification to avalanche, achieving tamper-proof and non-repudiation of the "replicable evidence snapshot".

[0087] S700. The hash chain evidence snapshot is used by the verification party to regenerate the standardized input summary for verification, re-execute the solution of the convex quadratic programming model, and verify the newly generated standardized input summary for verification, witness variable strength, KKT residual, address cluster merging result and hash chain integrity based on the candidate associated edge to be verified and the corresponding data in the evidence snapshot.

[0088] In this embodiment, verification refers to the verification party regenerating the standardized input summary for verification based on the candidate associated edges to be verified and the corresponding data in the evidence snapshot, re-executing the convex quadratic programming model corresponding to the candidate associated edges, and performing consistency verification on the standardized input summary for verification, the strength of witness variables, KKT residuals and address cluster merging results. At the same time, the continuity and integrity of the hash chain are verified to confirm that the input data, calculation process, numerical convergence status and evidence records corresponding to the inference results of the manipulating entity have not undergone unauthorized changes.

[0089] In some embodiments, the re-verification includes the following steps: The verification party queries the evidence snapshot corresponding to each candidate associated edge based on the candidate associated edge to be verified, and regenerates the verification standardized input summary of each candidate associated edge based on the candidate associated edge identifier, the address cluster identifiers at both ends of the candidate associated edge, and the standardized model input data (the generation steps of the standardized input summary are the same as S300); and compares the verification standardized input summary with the gated verification standardized input summary saved in the evidence snapshot; For candidate association edges to be re-verified, query the evidence snapshots corresponding to each candidate association edge, and re-solve the convex quadratic programming model corresponding to each candidate association edge based on the standardized model input data, preset importance weight vectors, and KKT residual calculation parameters. Calculate the witness variable strength and KKT residuals, and verify their consistency with the corresponding content saved in the evidence snapshots. After all candidate association edges pass the re-verification, reconstruct the association graph using these candidate association edges, re-cluster and merge address clusters, and generate a re-verification merging result summary, which is then compared with the merging result summary in the evidence snapshots.

[0090] Based on the previous hash pointer and the current hash pointer in the evidence snapshot corresponding to the candidate related edge to be re-verified, the continuity and integrity of the hash chain are verified sequentially. If any of the following occurs: the standardized input summary for re-verification is inconsistent with the standardized input summary for gated verification saved in the evidence snapshot; the strength of the witness variable, the KKT residual, and the corresponding content saved in the evidence snapshot are inconsistent within the preset numerical tolerance range; the merged result summary is inconsistent; or the hash chain verification fails, then the re-verification is marked as failed.

[0091] This step completes the closed loop of the entire methodology. It demonstrates how a third-party auditing firm, by combining the original input data to be verified with a hash chain evidence snapshot, technically delivers the tamper-proof and non-repudiable "verifiable" function through a zero-trust mechanism of "summary comparison, recalculation, and hash verification." This provides judicial evidence collection with a traceable, verifiable, and reverifiable computational evidence chain linked to the original input data, model calculation results, and hash chain evidence snapshot. This allows the inferences made by the manipulating entity to serve as the technical basis for subsequent evidence review, accountability tracking, and dispute verification.

[0092] In summary, this embodiment, by introducing an environmental disturbance subspace and orthogonal projection, isolates universal noise across the entire network from a mathematical and geometric perspective, reducing erroneous associations caused by macroscopic fluctuations in the blockchain.

[0093] By employing a convex quadratic programming model with slack variables, the inference of the manipulator is transformed into a mathematical optimization problem with a globally optimal solution and KKT residual metric. This provides objective, transparent, and mathematically measurable technical evidence for the inference of the manipulator that can be verified in audit or judicial scenarios.

[0094] By registering standardized input summaries before solving and performing ternary gating based on witness variable strength, KKT residuals, and input summary consistency, the risk of incorrect merging caused by input substitution or insufficient solution quality is reduced.

[0095] By using hash chain evidence snapshots to solidify standardized input summaries, key model parameters, witness variable strengths, and KKT residuals, a traceable and verifiable computational evidence chain is provided for third-party verification.

[0096] A cross-auditing closed loop of "pre-summary + ternary gating + hash chain" is adopted. This not only reduces I / O congestion caused by invalid records, but also ensures the integrity of the computational evidence chain and the irreversibility of the computation timeline, providing reliable technical support for third-party institutions to implement zero-trust independent verification.

[0097] Example 2: like Figure 2 As shown, this embodiment also provides a system for inferring the controlling entity and replicable data snapshots for public blockchain address clusters, including: The data standardization and noise reduction module is used to extract statistical values ​​of address clusters on a preset associated evidence dimension from blockchain ledger event and network propagation observation data, and splice them to generate the interaction feature vector of the current batch of address clusters; based on historical stable observation data, an environmental disturbance subspace and an orthogonal projection operator are constructed, and the interaction feature vector is projected onto the orthogonal complement space of the environmental disturbance subspace to obtain the environmental disturbance-free associated feature vector.

[0098] The optimization model construction module is used to obtain the observation matrix and observation target vector of candidate associated edges based on the associated feature vectors, set witness variables and slack variables, and construct the convex quadratic programming model corresponding to the candidate associated edges based on the two-sided inequality constraint of capital conservation and the feature identifiability inequality constraint.

[0099] The pre-solution registration module is used to perform deterministic encoding and hash calculation on the standardized input data corresponding to the candidate associated edges before solving the convex quadratic programming model, and to generate and register the standardized input summary before solving.

[0100] The execution and residual calculation module is used to solve the convex quadratic programming model corresponding to each candidate associated edge, obtain the optimal solution of the witness variable, calculate the strength of the witness variable by weighting according to the components of each witness variable in the optimal solution and their corresponding preset importance weights, and determine the maximum value of each KKT conditional residual corresponding to the solution result as the KKT residual.

[0101] The ternary gating module is used to perform ternary gating judgment on candidate associated edges based on the strength of the witness variable, the KKT residual, and the consistency between the standardized input summary of the gating check and the standardized input summary before the solution.

[0102] The output and snapshot generation module is used to uniformly construct an association graph for all candidate association edges that have passed the ternary gating judgment, perform clustering and merging on interconnected address clusters, output the inference results of the manipulator, and generate a hash chain evidence snapshot.

[0103] The verification module uses the hash chain evidence snapshot for the verification party to regenerate the standardized input summary for verification, re-execute the solution of the convex quadratic programming model, and verify the newly generated standardized input summary for verification, witness variable strength, KKT residual, address cluster merging result, and hash chain integrity based on the candidate associated edges to be verified and the corresponding data in the evidence snapshot.

[0104] It should be noted that the control subject inference and verifiable data snapshot system for public blockchain address clusters in this embodiment can be applied to the control subject inference and verifiable data snapshot method for public blockchain address clusters described in Embodiment 1, and the specific technical features involved in the method are consistent with those in Embodiment 1.

[0105] Example 3: This embodiment also provides an electronic device, including: One or more processors; One or more memories for storing one or more computer programs, and one or more processors for executing the one or more computer programs stored in the memories, so that the one or more processors perform the features / steps of a method for inferring the manipulator and taking verifiable data snapshots for a public blockchain address cluster as described in Embodiment 1.

[0106] Furthermore, at the hardware level, the device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other necessary hardware. One or more embodiments of this application can be implemented in software, for example, by the processor reading the corresponding computer program from the non-volatile memory into memory and then running it.

[0107] Of course, in addition to the software implementation, one or more embodiments of this application do not exclude other implementation methods, such as logic devices or a combination of hardware and software. That is to say, the execution subject of the above processing flow is not limited to each logic module, but can also be hardware or logic devices.

[0108] Example 4: Those skilled in the art will understand that all or part of the features / steps of the above-described method embodiments can be implemented by methods, data processing systems, or computer programs. These features may be implemented without hardware, entirely in software, or in a combination of hardware and software. The aforementioned computer program may be stored in one or more computer-readable storage media. When the computer program is executed (e.g., by a processor), it performs the steps of the above-described method embodiment for inferring the controlling entity and taking verifiable data snapshots for public blockchain address clusters.

[0109] The aforementioned storage media capable of storing program code include: solid-state drives, random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), optical storage devices, magnetic storage devices, flash memory, magnetic disks or optical disks and / or combinations thereof, that is, they can be implemented by any type of volatile or non-volatile storage devices or combinations thereof.

[0110] Example 5: This embodiment provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the steps of a method for inferring the controlling entity and taking verifiable data snapshots for a public blockchain address cluster as described in Embodiment 1.

[0111] It should be understood that although the steps in the flowcharts of the above embodiments 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 processes of the above embodiments 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.

[0112] The above description is merely a preferred embodiment of the present invention. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the present invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for inferring the controlling entity and taking verifiable data snapshots for public blockchain address clusters, characterized in that, include: Statistical values ​​of address clusters on a preset associated evidence dimension are extracted from blockchain ledger event and network propagation observation data, and the interaction feature vectors of the current batch of address clusters are spliced ​​together; an environmental disturbance subspace and an orthogonal projection operator are constructed based on historical stable observation data, and the interaction feature vectors are projected onto the orthogonal complement space of the environmental disturbance subspace to obtain the environmental disturbance-free associated feature vectors. The observation matrix and observation target vector of the candidate associated edge are obtained based on the associated feature vector. Witness variables and slack variables are set, and the convex quadratic programming model corresponding to the candidate associated edge is constructed based on the two-sided inequality constraint of capital conservation and the feature identifiability inequality constraint. Before solving the convex quadratic programming model, the standardized input data corresponding to the candidate associated edges are deterministically encoded and hashed to generate and register the standardized input summary before solving. Solve the convex quadratic programming model corresponding to each candidate association edge to obtain the optimal solution of the witness variable. Calculate the strength of the witness variable based on the components of each witness variable and their corresponding importance weights in the optimal solution. Calculate the KKT residual based on the solution results. Based on the strength of the witness variable, the KKT residual, and the consistency between the standardized input summary of the gate check and the standardized input summary before the solution, a ternary gate check is performed on the candidate associated edges. A unified association graph is constructed for all candidate association edges that pass the ternary gating judgment. Clustering and merging are performed on interconnected address clusters, the inference results of the manipulator are output, and a hash chain evidence snapshot is generated. The hash chain evidence snapshot is used by the verification party to regenerate the standardized input summary for verification, re-execute the solution of the convex quadratic programming model, and verify the newly generated standardized input summary for verification, witness variable strength, KKT residual, address cluster merging result, and hash chain integrity based on the candidate associated edge to be verified and the corresponding data in the evidence snapshot. Constructing the convex quadratic programming model corresponding to the candidate associated edges includes: The objective function of the convex quadratic programming model includes an error penalty term calculated based on the observation matrix of the candidate associated edges and the observation target vector, a regularization smoothing term for the witness variables, and linear penalty terms corresponding to the capital conservation deviation slack variables and the feature identifiability slack variables. The constraints of the convex quadratic programming model include at least two-sided inequality constraints on capital conservation and identifiability constraints. The two-sided inequality constraints on capital conservation consist of a capital conservation deviation constraint matrix, a constraint objective vector, witness variables, a capital conservation error bound vector, and capital conservation deviation slack variables. The identifiability constraints on identifiability consist of a identifiability constraint matrix, a identifiability lower bound vector, witness variables, and identifiability slack variables. Output the manipulation subject inference results for candidate related edges determined by ternary gating and generate a hash chain evidence snapshot, including: Using the address clusters at both ends of the candidate association edges determined by the ternary gating as nodes, and the candidate association edges determined by the ternary gating as the connection relationship, an association graph is constructed. Clustering and merging are performed on the interconnected address clusters in the association graph to obtain merged address clusters. A merging result summary is generated based on the merged address clusters. The merged address clusters are used to represent the inference results of the manipulation subject. The candidate associated edge identifier, the address cluster identifiers at both ends of the candidate associated edge, the standardized input summary of the gated verification, the witness variable strength, the KKT residual, and the merged result summary are serialized according to a preset deterministic coding rule to obtain the snapshot payload. The standardized model input data used to construct the convex quadratic programming model corresponding to the candidate associated edge, the preset importance weight vector, the KKT residual calculation parameters, and the preset numerical tolerance used during the re-verification are serialized according to a preset deterministic coding rule to obtain the snapshot header metadata. The hash value of the preceding hash pointer, the hash value of the snapshot payload, and the hash value of the snapshot header metadata are concatenated and the hash is calculated again to generate the current hash pointer, which serves as the preceding hash pointer for the next evidence snapshot. The snapshot payload, the snapshot header metadata, and the hash pointer together constitute the hash chain evidence snapshot.

2. The method for inferring the controlling entity and taking verifiable data snapshots for public blockchain address clusters according to claim 1, characterized in that, The step of constructing an environmental disturbance subspace and an orthogonal projection operator based on historical stable observation data, and projecting the interactive feature vector onto the orthogonal complement space of the environmental disturbance subspace to obtain the environmental disturbance-free associated feature vector includes: Singular value decomposition is performed on the matrix corresponding to the acquired historical stable observation data to obtain the left singular vector matrix, the singular value diagonal matrix, and the right singular vector matrix; The cutoff parameter r is determined based on the preset cumulative energy threshold, and the first r left singular vectors are selected to form a matrix. and the matrix The column space is used as the environmental perturbation subspace, and an orthogonal projection operator is constructed based on the environmental perturbation subspace; Subtract the historical stable observation mean vector from the interaction feature vector of the current batch address cluster, and then multiply it by the orthogonal projection operator to obtain the environmentally disturbed associated feature vector.

3. The method for inferring the controlling entity and taking verifiable data snapshots for public blockchain address clusters according to claim 1, characterized in that, The witness variable includes multiple components, which correspond to multiple association evidence dimensions enabled by the current convex quadratic programming model. The dimensions of the associated evidence include the common interaction dimension, the change pattern dimension, the broadcast time sequence dimension, and the amount distribution dimension; the statistical value of the common interaction dimension is used to characterize the similarity of two address clusters in the same batch or adjacent batches when they jointly input, jointly participate in transactions, jointly aggregate, or jointly split; the statistical value of the change pattern dimension is used to characterize the similarity of two address clusters in terms of change address generation rules, change ratio, or change address usage patterns. The statistical values ​​of the broadcast time sequence dimension are used to characterize the similarity of the first observation time, arrival time difference, or propagation path of related transactions during network propagation; the statistical values ​​of the amount distribution dimension are used to characterize the similarity of the distribution of inflow amount, outflow amount, net flow amount, or amount quantile between two address clusters. Each component in the optimal solution of the witness variable is used to characterize the degree of evidence support for the inference that the address clusters at both ends of the candidate associated edge are controlled by the same manipulator, according to the corresponding associated evidence dimension.

4. The method for inferring the controlling entity and taking verifiable data snapshots for public blockchain address clusters according to claim 1, characterized in that, The generation and registration of the standardized input summary before solving includes: According to the preset deterministic coding rules, the standardized model input data and candidate associated edge basic data used to construct the convex quadratic programming model corresponding to the candidate associated edge are encoded and hashed to obtain the standardized input summary before the solution; the standardized input summary before the solution is recorded using the candidate associated edge identifier as an index.

5. The method for inferring the controlling entity and taking verifiable data snapshots for public blockchain address clusters according to claim 1, characterized in that, Solve the convex quadratic programming model corresponding to each candidate association edge, calculate the strength of the witness variables, and calculate the KKT residuals based on the solution results, including: Solve the convex quadratic programming model corresponding to the candidate association edge to obtain the optimal solution of the witness variable vector, wherein the components of the witness variable vector are arranged in a fixed order according to the association evidence dimension; Based on the optimal solution of the witness variable vector and the preset importance weight vector consistent with the fixed order of the associated evidence dimensions, the witness variable strength of the candidate associated edge is calculated to characterize the comprehensive degree of association between the address clusters at both ends of the candidate associated edge under the control of the same manipulating entity. The two-sided inequality constraint of capital conservation, the identifiability inequality constraint, and the variable range constraint are equivalently transformed into inequality constraints in a unified direction, and the KKT residual is calculated based on the solution results. The KKT residual is determined by the maximum residual among the original feasibility residual, the dual feasibility residual, the complementary relaxation residual, and the dual variable nonnegativity residual, and is used to characterize the numerical convergence accuracy of the convex quadratic programming model corresponding to the candidate associated edge.

6. The method for inferring the controlling entity and taking verifiable data snapshots for public blockchain address clusters according to claim 1, characterized in that, The determination condition corresponding to the three-element gating system must simultaneously satisfy the following three conditions: The strength of the witness variable is greater than or equal to the preset association strength threshold; The KKT residual is less than or equal to a preset residual threshold; The gating verification normalized input summary generated from the current input is consistent with the normalized input summary before solving.

7. A method for inferring the controlling entity and taking verifiable data snapshots for public blockchain address clusters according to any one of claims 1 to 6, characterized in that, This also includes re-verification following the steps below: The verification party regenerates the standardized input summary for verification based on the candidate associated edges to be verified and the corresponding data in the evidence snapshot, re-executes the solution of the convex quadratic programming model corresponding to the candidate associated edges, and performs consistency verification on the standardized input summary for verification, the strength of witness variables, KKT residuals and address cluster merging results. At the same time, it verifies the continuity and integrity of the hash chain to confirm that the input data, calculation process, numerical convergence status and evidence records corresponding to the inference results of the manipulating entity have not undergone unauthorized changes.

8. An electronic device, characterized in that, include: At least one processor; At least one memory used to store processor-executable instructions; The at least one processor implements a method for inferring the controlling entity and taking verifiable data snapshots for a public blockchain address cluster as described in any one of claims 1 to 7 by running the executable instructions.

Images