Methods, apparatus, devices and storage media for evaluating incentives for participants in privacy-preserving computation
By acquiring the computing, network, and data resource parameters of the participants in privacy-preserving computation, and combining them with the data implementation costs and model performance contributions, the comprehensive score and excess revenue of each participant are quantified. This solves the problem of unfair evaluation in existing technologies and achieves a fairer distribution of revenue and more efficient computation.
Patent Information
- Application Number
- CN202611202725.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-07
- Publication Date
- 2026-11-10
AI Technical Summary
Existing privacy-preserving computing collaborations struggle to comprehensively assess the resource input and contributions of each participant, leading to discrepancies between assessment results and actual value. This is especially true in multi-party collaboration scenarios where computational overhead is high and allocation is unfair.
By acquiring the computing, network, and data resource parameters of each participant, and combining them with the data implementation cost and model performance contribution, the comprehensive score and excess revenue of each participant are quantified, and their revenue share and incentive value are determined.
This enables a more comprehensive and reasonable quantification of the contributions of all participants, improves the fairness of revenue distribution and computational efficiency, ensures the enthusiasm of high-quality data and high-resource contributors, and stabilizes multi-party privacy computing cooperation.
Smart Images

Figure CN123071016A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of privacy computing, and in particular to a method, apparatus, device, and storage medium for evaluating incentives for participants in privacy computing. Background Technology
[0002] When multiple institutions collaborate on privacy-preserving computation, they typically use methods such as federated learning and multi-party secure computation to jointly model the data, and evaluate the contributions of the participants and distribute the benefits based on existing rules.
[0003] However, existing evaluation schemes often focus on the contribution of data to the model's performance, or allocate data based on data volume and historical input, making it difficult to simultaneously reflect actual resource inputs such as data quality, computing power consumption, and network transmission. In scenarios with more participants and cross-platform collaboration, problems such as high computational overhead and deviations between evaluation results and actual value are also likely to occur.
[0004] Therefore, how to balance comprehensive evaluation, fair allocation, and computational efficiency in multi-party privacy computing collaborations, and accurately quantify the actual contributions of each participant, has become a technical problem that needs to be solved. Summary of the Invention
[0005] This application provides a method for evaluating incentives for participants in privacy-preserving computation to address the aforementioned technical problems. This method addresses the needs of multi-participant contribution evaluation and benefit distribution in privacy-preserving computation collaboration scenarios. It comprehensively considers the resource input, data value, implementation costs, and contributions to model performance improvement made by each participant during the collaboration process. It quantifies the actual contribution of each participant and, based on this, determines a benefit distribution result and incentive value matching their contribution, thereby balancing the comprehensiveness of the evaluation, the fairness of the distribution, and the rationality of the evaluation results.
[0006] In a first aspect, embodiments of this application provide a method for evaluating incentives for participants in privacy computing, comprising: acquiring, for each participant in a privacy computing collaboration, information on the participant's use of computing resources, network resources, and data resources; the data resources parameters include the number of data entries, the number of feature dimensions, the update frequency, and data quality indicators; determining the comprehensive score of each participant based on the participant's information on the use of computing resources, network resources, and data resources; acquiring the data implementation cost and model performance contribution value of each participant; wherein, the data implementation cost is used to characterize the total implementation cost invested by the participant in participating in privacy computing, and the model performance contribution value is used to characterize the marginal improvement in model performance brought about by the participant's participation; calculating the excess revenue based on the model performance contribution value and data implementation cost of each participant in the privacy computing collaboration; determining the revenue share of each participant based on the comprehensive score and model performance contribution value of each participant in the privacy computing collaboration; and determining the incentive value of each participant based on the data implementation cost of each participant and the product of the revenue share and the excess revenue of each participant.
[0007] In one possible embodiment, a participant's comprehensive score is determined based on the participant's computing resource usage information, network resource usage information, and data resource parameters. This includes: determining a data resource score based on the number of data entries, the number of feature dimensions, and the update frequency; determining a data quality score based on data quality indicators; determining a participant's computing resource score based on the participant's computing resource usage information; determining a participant's network resource score based on the participant's network resource usage information; and determining a participant's comprehensive score based on the participant's data resource score, data quality score, computing resource score, and network resource score.
[0008] In one possible embodiment, the comprehensive score of a participant is determined based on the participant's data resource score, data quality score, computing resource score, and network resource score. This includes: determining the weight coefficients for each of the data resource score, data quality score, computing resource score, and network resource score based on the cooperation scenario of privacy computing cooperation; and weighting the data resource score, data quality score, computing resource score, and network resource score based on the determined weight coefficients to obtain the comprehensive score of the participant.
[0009] In one possible embodiment, the data implementation cost of each participant is obtained, including: for each participant, if the participant is a partner, the data implementation cost of the partner is determined based on the partner's data acquisition cost, data cleaning and de-identification cost, storage space rental cost, security guarantee cost, and daily operation and maintenance and disaster recovery cost.
[0010] In one possible embodiment, the data implementation cost of each participant is obtained, including: for each participant, if the participant is a relay party, the data implementation cost of the relay party is determined based on the relay party's security aggregation coordination and computing power cost, security assurance cost, and daily operation and maintenance and disaster recovery cost.
[0011] In one possible embodiment, the method further includes: retrieving, for each participant, supporting materials that have undergone hashing and encryption from a blockchain or a secret shared distributed storage; decrypting the supporting materials and verifying the decrypted supporting materials based on the hash value; if the verification passes, verifying the data implementation cost of the participants based on the decrypted supporting materials, and calculating excess revenue and contribution value based on the verified data implementation cost.
[0012] In one possible embodiment, the method further includes: if the verification fails, adjusting the data implementation cost of the participating party based on the decrypted supporting materials, and calculating the excess revenue and contribution value based on the adjusted data implementation cost.
[0013] In one possible embodiment, obtaining the model performance contribution value of each participant includes: obtaining the model performance index of each participant subset, wherein the participant subset consists of some participants in the privacy computing collaboration; if the number of participants in the privacy computing collaboration is less than or equal to a preset number, then for each participant, based on the Shapley value formula and the model performance index of each target participant subset excluding the participant, calculating the expected value of the marginal improvement in model performance after the participant joins each target participant subset, thereby obtaining the model performance contribution value of the participant.
[0014] In one possible embodiment, the method further includes: if the number of participants in the privacy computing collaboration is greater than a preset number, then performing a random permutation operation on all participants in the privacy computing collaboration T times in a loop; after each random permutation operation, extracting a temporary subset formed by all participants before the current participant in the current permutation, obtaining the model performance index of the temporary subset and the model performance index after the current participant is added to the temporary subset; calculating the difference in model performance index before and after the current participant is added to the temporary subset to obtain the marginal improvement in model performance of the current participant under the current permutation; and calculating the average of the marginal improvement in model performance of the current participant under each permutation obtained by the T random permutation operations to obtain the model performance contribution value of the current participant.
[0015] In one possible embodiment, the revenue share of each participant is determined based on the overall score and model performance contribution value of each participant in the privacy computing collaboration. This includes: for each participant, determining the participant's revenue share as the ratio of the product of the participant's overall score and model performance contribution value to the cumulative value of the product of the overall score and model performance contribution value of each participant.
[0016] In one possible embodiment, the incentive value for each participant is determined based on the data implementation cost of each participant and the product of the participant's revenue share and excess revenue. This includes determining the incentive value for each participant as the sum of the product of the participant's data implementation cost and the participant's revenue share and excess revenue.
[0017] In one possible embodiment, obtaining computing resource usage information, network resource usage information, and data resource parameters of participating parties in privacy computing collaboration includes: monitoring the resource consumption of each participating party's computing nodes, collecting computing power consumption time, memory usage capacity, and storage space capacity to obtain computing resource usage information of each participating party; monitoring the transmission link and traffic operation indicators of each participating party, collecting the total number of transmitted bytes, the number of routed and forwarded data packets, security aggregation computing power consumption, and data packet loss rate to obtain network resource usage information; and monitoring data flow resources through data resource probes, collecting the number of data entries, feature dimensions, update frequency, and data quality indicators to obtain data resource parameters.
[0018] Secondly, embodiments of this application provide a privacy computing participant incentive evaluation device, comprising: a data acquisition module, used to acquire, for each participant in a privacy computing collaboration, information on the participant's computing resource usage, network resource usage, and data resource parameters; the data resource parameters include the number of data entries, the number of feature dimensions, the update frequency, and data quality indicators; a comprehensive score determination module, used to determine the comprehensive score of each participant based on the participant's computing resource usage, network resource usage, and data resource parameters; and a cost acquisition module, used to acquire the data implementation cost and model performance contribution value of each participant; wherein, the data... The implementation cost is used to characterize the total implementation cost invested by the participants in participating in privacy computing. The model performance contribution value is used to characterize the marginal improvement in model performance brought about by the participation of the participants. The excess revenue calculation module is used to calculate the excess revenue based on the model performance contribution value and data implementation cost of each participant in the privacy computing collaboration. The revenue share determination module is used to determine the revenue share of each participant based on the comprehensive score and model performance contribution value of each participant in the privacy computing collaboration. The incentive value determination module is used to determine the incentive value of each participant based on the data implementation cost of each participant and the product of the revenue share and excess revenue of each participant.
[0019] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory, causing the processor to perform the method provided above.
[0020] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided above.
[0021] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method described above.
[0022] The privacy computing participant incentive evaluation method, apparatus, device, and storage medium provided in this application obtain computing resource usage information, network resource usage information, and data resource parameters of each participant in privacy computing cooperation, determine the comprehensive score of each participant, and combine the data implementation cost, model performance contribution value, and the excess benefit, benefit ratio, and incentive value calculated therefrom for each participant. This enables a coordinated quantitative evaluation of resource input, data element value, implementation cost, and model performance gain, thereby more accurately reflecting the actual contribution of each participant in privacy computing collaboration and improving the fairness, rationality, and applicability of benefit distribution and incentive results. Attached Figure Description
[0023] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0024] Figure 1 A flowchart illustrating the privacy computing participant incentive assessment method provided for this application;
[0025] Figure 2 A framework diagram of a privacy computing collaboration system provided for this application;
[0026] Figure 3 A schematic diagram of the privacy computing participant incentive evaluation device provided in this application;
[0027] Figure 4 A schematic diagram of the structure of the electronic device provided in this application.
[0028] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation
[0029] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application.
[0030] Privacy-preserving computation collaboration and revenue sharing fall under the fields of secure data circulation and collaborative modeling, and are commonly found in multi-party data collaboration scenarios such as financial risk control, medical collaborative analysis, and government data sharing. In a typical architecture, institutions such as banks, telecom operators, and e-commerce platforms each hold local data and conduct federated learning or multi-party secure computation through their respective privacy-preserving computation platforms to complete model training and result collaboration without the data leaving its domain.
[0031] In existing approaches, after joint modeling is completed, multiple organizations typically evaluate the contributions of each participant according to pre-defined rules and allocate the benefits of collaboration accordingly. Common practices include measuring contributions based on data volume, historical input, or the improvement in metrics such as model accuracy and AUC (Area Under the ROC Curve) after a participant joins. The basic idea is to estimate the value of participation from a single dimension, such as data contribution or cost input.
[0032] However, such solutions fail to accurately reflect the actual input in multi-party collaboration. On the one hand, a large number of data entries does not equate to high data value; differences in feature dimensions, update frequency, and data quality all affect the effectiveness of joint modeling. On the other hand, computational resource consumption and network transmission costs are also key ongoing inputs during the collaboration process. If these are not included in a unified assessment, it can easily lead to a mismatch between the distribution of benefits and the actual contributions.
[0033] In cross-institutional and cross-platform collaboration scenarios, as the number of participants increases, evaluating solely based on model performance contributions can impose a significant computational burden and may lead to results that deviate from their true value due to overly narrow evaluation dimensions. Consequently, the enthusiasm of high-quality data providers and those with high resource investment is dampened, while low-quality or low-investment participants may receive disproportionate benefits, thus affecting the stability and sustainability of privacy-preserving computing collaborations.
[0034] In light of this, balancing comprehensive evaluation, fair allocation, and computational efficiency in multi-party privacy-preserving computation collaborations has become a pressing technical challenge. To address this issue, a method for evaluating incentives for participants in privacy-preserving computation is proposed. This method obtains information on each participant's computing resource usage, network resource usage, and data resource parameters during the collaboration. Based on this information, a comprehensive score is determined for each participant. Furthermore, by combining each participant's data implementation costs, model performance contributions, and the calculated excess returns, the percentage of returns and incentive values are further determined, thus providing a more comprehensive and reasonable quantification of each participant's actual contribution. This method is applicable to collaborative architectures where multiple institutions conduct joint modeling through a privacy-preserving computation platform, providing a unified basis for subsequent incentive allocation.
[0035] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.
[0036] Figure 1 A flowchart illustrating the privacy computing participant incentive assessment method provided in this application is shown below. Figure 1 As shown, the method includes:
[0037] S101: For each participant in the privacy computing collaboration, obtain information on the participants' use of computing resources, network resources, and data resources in the privacy computing collaboration.
[0038] Privacy-preserving computation, in essence, involves joint modeling or statistical analysis under the premise that data remains within its domain and raw data is not shared. Participants in privacy-preserving computation collaborations can be categorized into collaborators and intermediaries based on their functions. Collaborators are the data subjects, holding the original business data and possessing data ownership. They also have joint business needs and are the value drivers of the privacy-preserving computation task. Intermediaries, on the other hand, do not hold the original business data and only provide services such as scheduling, computing power, network forwarding, encrypted transmission, and trusted coordination; they do not participate in reading or writing the original business data.
[0039] Computational resource usage information characterizes the actual computing resources used by participants during privacy-preserving computation collaboration, such as computation time consumption and memory usage. Network resource usage information characterizes the network communication resources used by participants during collaboration, such as the number of bytes transmitted, bandwidth usage, and latency. Data resource parameters characterize the features of the data resources provided by participants, including the number of data entries, the number of feature dimensions, update frequency, and data quality indicators. Specifically, the number of data entries represents the total number of samples used by the participants for privacy-preserving computation; the number of feature dimensions represents the number of feature fields in the samples provided by the participants; the update frequency represents the update cycle of the participants' samples or datasets; and data quality indicators include, but are not limited to, missing value rate, outlier percentage, label completeness, and data compliance and anonymization.
[0040] In practice, the resource usage and data statistics of participating parties during task execution can be collected to obtain information on each party's computing resource usage, network resource usage, and data resource parameters. For partners in privacy-preserving computing collaborations, each partner can collect and report their own computing resource usage, network resource usage, and data resource parameters. For intermediaries, the intermediaries can collect their own computing resource usage and network resource usage information. After each participating party completes its collection, it can report to an evaluation service node, such as one of the intermediaries or other institutions. This evaluation service node will then execute the incentive evaluation method provided in this application to determine the incentive value for each participating party.
[0041] For example, all parties involved in privacy-preserving computation, such as organizations, can collaborate on data based on the data circulation components built into their own privacy-preserving computation platforms. The transmission networking module within the data circulation component is the core of multi-party resource networking. It assigns a unique network identifier (NodeID) to each participating party upon registration. During registration, participating parties must submit organizational information, platform information, and platform-supported resource information such as computing power and memory, providing a foundation for subsequent multi-party collaboration and incentive value calculation. The transmission networking module can dynamically connect different projects and network segments according to the needs of the collaboration scenario, improving the efficiency of multi-party data collaboration. The transmission networking module can also configure resource probe strategies as needed to conduct unified statistical verification of the actual usage of registered resources by each party during the collaboration process. The aforementioned three types of resource information—computing resource usage information, network resource usage information, and data resource parameters—are all collected using the supporting capabilities of the data circulation component.
[0042] The computing resource usage information is used to characterize the actual computing power and storage hardware resources used by the participating parties during the collaboration process. It covers metrics such as computing power consumption time and memory usage. Real-time usage data for CPU, GPU, memory, and storage can be collected through resource monitoring probes deployed on each node, or partners can proactively report resource usage. Network resource usage information characterizes the network communication resources used by each party during the collaboration, including metrics such as the number of bytes transmitted, bandwidth usage, and network latency. For directly connected nodes, the transmission networking module collects interaction traffic and latency. Nodes performing forwarding functions additionally use the routing forwarding module to calculate transit traffic overhead. Only encrypted communication data is collected throughout the process; the original plaintext business data is not touched. Data resource parameters are only for partners holding data and are used to characterize various features of the datasets they provide externally, including the number of data entries, the number of feature dimensions, update frequency, and multi-dimensional data quality indicators. This type of information is automatically collected through the partner's local standardized API interface. Data quality indicators are further subdivided into completeness representing the proportion of missing values, accuracy representing the proportion of outliers, and consistency representing the cross-source data conflict rate.
[0043] After the three types of resource information—computing resource usage information, network resource usage information, and data resource parameters—are collected, they will be uniformly handed over to one of the participants' incentive value evaluation and measurement modules to conduct multi-dimensional resource evaluation and cost and contribution fusion measurement, and finally obtain the incentive value of each participant.
[0044] S102: Determine the comprehensive score of each participant based on their computing resource usage information, network resource usage information, and data resource parameters.
[0045] The overall score is used to uniformly represent the resource contributions of the participants in privacy computing collaboration. Its essence is to map the input of computing resources, network resources, and data resources to the same dimension to form comparable results.
[0046] Because the input indicators differ in units, numerical ranges, and degree of impact on cooperation value, the information on computing resource usage, network resource usage, and data resource parameters can be standardized before determining the overall score to reduce the impact of differences in the units of measurement between different indicators. For example, normalization, interval mapping, or other suitable processing methods can be used to convert the relevant indicators into results that are easy to compare.
[0047] After processing, a comprehensive score is obtained based on the participants' computing resource usage information, network resource usage information, and data resource parameters. In practice, consistency checks and missing information checks can be performed on the acquired information, and supplementation or re-acquisition can be performed when necessary to improve the reliability of the comprehensive score result.
[0048] It can calculate a score for various parameters in resource usage information, network resource usage information, and data resource parameters, and then weight the scores corresponding to multiple parameter categories to obtain a comprehensive score. When weighting, either fixed weights or dynamically changing weights can be used.
[0049] The scoring logic for sub-indicators within a single resource category is as follows: Taking computing resources as an example, which include multiple sub-parameters such as computing power duration, memory usage, and storage space, all sub-indicators are first normalized to eliminate differences in units and magnitudes, uniformly mapping them to a score range of 0-100. Then, internal sub-weights are set for sub-indicators such as computing power, memory, and storage, and the computing resource dimension score is calculated by weighting these weights. Network resources and data resources use the same internal scoring logic. Network resources are weighted by sub-indicators such as transmission bytes, bandwidth usage, and transmission latency to obtain the network dimension score. Data resources are weighted by sub-indicators such as the number of data entries, the number of feature dimensions, update frequency, data integrity, accuracy, and consistency to obtain the data dimension score. The computing resource dimension score, data dimension score, and network dimension score are then weighted to obtain the comprehensive score.
[0050] This step processes inputs from different resource dimensions in a unified manner to obtain a comprehensive score that can be directly used for subsequent revenue distribution, thereby enabling the resource contributions of the participants to be included in the evaluation.
[0051] S103: Obtain the data implementation costs and model performance contributions of each participant.
[0052] Among them, data implementation cost is used to characterize the total implementation cost invested by the participants in participating in privacy computing, and model performance contribution value is used to characterize the marginal improvement in model performance brought about by the participation of the participants.
[0053] Data implementation cost quantifies the actual costs incurred by participating parties to enable their data to be integrated into the privacy computing collaboration, such as data preprocessing costs, anonymization costs, dataset maintenance costs, and encryption costs. Model performance contribution value quantifies the marginal improvement in model performance brought about by a participating party joining the privacy computing collaboration. Model performance metrics can be selected according to the task type, such as accuracy, recall, AUC, F1 score, and positive / negative sample discrimination, and are used to characterize the marginal improvement in model performance brought about by the participating party.
[0054] In practice, cost information submitted by participating parties can be categorized, deduplicated, and aligned with the timeframe. Investments unrelated to the current privacy computing task can be eliminated, retaining only the calculable costs corresponding to the current collaboration, thereby obtaining the data implementation cost within the task period.
[0055] The model performance contribution value can be determined based on the changes in model performance before and after the participation of the participants. The Shapley value formula can be used to calculate the model performance contribution.
[0056] This step extracts the actual role of the participants in the cooperation from the perspectives of cost input and model marginal value, establishing a quantifiable basis for subsequent collection of excess returns and calculation of return ratio.
[0057] S104: Calculate excess revenue based on the model performance contribution and data implementation cost of each participant in the privacy computing collaboration.
[0058] In this step, excess revenue is used to characterize the distributable residual revenue generated by the privacy-preserving computing collaboration based on the joint modeling or joint analysis results after compensating each participant's task implementation costs. This residual revenue is not simply the total revenue of the collaboration, but rather the net added value tied to the participants' costs and model contributions.
[0059] The total revenue can be obtained by multiplying the sum of the model performance contributions of each participant by a preset performance conversion coefficient. The difference between this total revenue and the cumulative data implementation costs of each participant can then be calculated to obtain the excess revenue. The preset performance conversion coefficient is a pre-agreed quantitative conversion standard used to characterize the revenue corresponding to each 1 percentage point increase in model performance contribution.
[0060] S105: Determine the revenue share of each participant based on the comprehensive score and model performance contribution value of each participant in the privacy computing collaboration.
[0061] Among them, the revenue share is used to describe the relative share of each participant in the distribution of excess revenue. Its calculation takes into account both the comprehensive score on the resource input side and the model performance contribution value on the output side.
[0062] In practice, the evaluation service node first reads the comprehensive score generated in S102 and the model performance contribution value generated in S103, and performs interval consistency processing on the two types of indicators. If the comprehensive score and the model performance contribution value have unified dimensions, they can directly proceed to the combination calculation; if the distributions of the two differ significantly, they are normalized again. Then, the combination value of each participant is determined based on the comprehensive score and the model performance contribution value, and the revenue share of each participant is determined based on the combination value of each participant. This combination value can be the product, average, sum, or other value of the comprehensive score and the model performance contribution value, and the combination value is positively correlated with both the comprehensive score and the model performance contribution value.
[0063] This step couples the two dimensions of resource input and model performance improvement into a single allocation coefficient, so that the proportion of benefits is no longer determined by a single model indicator, nor does it rely solely on cost or data volume statistics, thus aligning with the actual contribution structure in multi-party collaboration.
[0064] S106: Determine the incentive value for each participant based on the data implementation cost of each participant and the product of the participant's share of revenue and excess revenue.
[0065] The incentive value ultimately represents the benefits each participant receives in the privacy computing collaboration and serves as the direct basis for settlement, reconciliation, or the distribution of points within the platform. Its calculation logic consists of two parts: one part is the compensation for the data implementation costs of the participating parties, and the other part is the excess revenue allocated according to the proportion of benefits received.
[0066] In practice, the implementation cost, excess revenue, and revenue share of data reading by service nodes are evaluated, according to the formula: Incentive Value = Number according to Calculate the incentive value for the corresponding participant by adding the implementation cost plus the percentage of revenue multiplied by the excess revenue.
[0067] In the specific execution process, the evaluation service node first generates a draft settlement record for each participant. This record includes the participant identifier, data implementation cost, revenue share, excess revenue, excess allocation amount, and incentive value. Then, boundary checks are performed, such as verifying whether the incentive value is non-negative and whether the sum of the incentive values for all participants equals the sum of the total implementation cost and excess revenue. The total implementation cost is the sum of the data implementation costs of all participants. If there are minor errors caused by floating-point operations, these errors can be adjusted to the record of the participant with the highest revenue share or the largest absolute value to ensure overall ledger balance.
[0068] After verification, the evaluation service node writes the incentive value into the settlement result table and sends an allocation instruction to the fund settlement system, contract management system, or platform account system. For scenarios where the incentive is distributed in monetary form, the fund settlement system executes the allocation according to the incentive value; for scenarios where the incentive is in the form of platform points or resource quotas, corresponding points or computing power vouchers can be generated according to the incentive value and exchange rules.
[0069] In one possible implementation, if the cooperation agreement stipulates a minimum replenishment ratio, a settlement cap, or a reserved risk reserve, the evaluation service node adds a rule verification module after formula calculation to perform rule-based correction on the incentive value, and simultaneously retains both the original calculated value and the corrected result to meet audit requirements. For example, when the settlement system requires two decimal places, the incentive value can be uniformly rounded during the final output stage.
[0070] This step combines cost recovery and excess profit distribution into the same settlement expression to form a directly executable incentive value for participants, enabling input compensation and contribution rewards in multi-party privacy computing collaborations to complete a closed loop under the same rules.
[0071] The privacy computing participant incentive evaluation method in this application establishes a unified incentive evaluation chain for multi-party privacy computing collaboration by incorporating computing power consumption, network usage, data scale, number of feature dimensions, update frequency, and data quality into a unified resource evaluation. It also combines end-to-end implementation costs with marginal improvements in model performance to collect and distribute excess benefits. Compared to methods that allocate incentives solely based on data volume, historical investment, or the magnitude of a single model improvement, this method maintains consistent evaluation criteria and an executable settlement process, quantifying high-quality data investment, significant computing power and network investment, and substantial model gains in the incentive value.
[0072] In one possible implementation, the comprehensive score of a participant is determined based on the participant's computing resource usage information, network resource usage information, and data resource parameters. This includes: determining a data resource score based on the number of data entries, the number of feature dimensions, and the update frequency; determining a data quality score based on data quality indicators; determining a participant's computing resource score based on the participant's computing resource usage information; determining a participant's network resource score based on the participant's network resource usage information; and determining the participant's comprehensive score based on the participant's data resource score, data quality score, computing resource score, and network resource score.
[0073] First, the number of data entries, the number of feature dimensions, and the update frequency are normalized, and then weighted according to preset weights to obtain a data resource score. Data quality indicators can be mapped to quality scores based on missing rate, noise rate, duplication rate, or validation pass rate. Computational resource usage information can be formed into resource scores based on training duration, peak computing power, average memory usage, and persistent storage usage. Network resource usage information can be formed into network scores based on upload and download traffic, number of routing and forwarding data packets, and packet loss rate. Subsequently, each score is standardized and then weighted and summed according to uniform weights to obtain the comprehensive score of the participants.
[0074] For example, the first Overall score of each participant The calculation relationship can be:
[0075]
[0076] in, , , and These are the weight coefficients for the scores of the corresponding items. They can be fixed values or determined based on the privacy computing scenario. The sum of these four weights is 1. , , and They represent the first Each participant's score is calculated based on its data resource score, computing resource score, network resource score, and data quality score.
[0077] For example, the data resource score can be the product of the number of data entries, the number of feature dimensions, and the update frequency, or the standardized values of these three factors. The computing resource score can be the product of computing power consumption and memory usage, or the standardized values of these two factors. The network resource score can be the product of the number of bytes transmitted, bandwidth usage, and latency, or the standardized values of these three factors.
[0078] For example, the data quality score can be a weighted result of each item in the data quality indicators. Taking data quality indicators including completeness, accuracy, and consistency as an example, the data quality score can be: 0.3 × Completeness + 0.4 × Preparedness + 0.3 × Consistency.
[0079] This comprehensive score can simultaneously reflect the participants' data scale, data quality, and computational and network inputs, providing a unified quantitative basis for subsequent revenue sharing and incentive value calculations. By incorporating multiple types of resource information into the same scoring system, the evaluation of participants' contributions can be more consistent with their actual inputs, thereby improving the matching and comparability of incentive allocation.
[0080] In one possible implementation, the comprehensive score of a participant is determined based on the participant's data resource score, data quality score, computing resource score, and network resource score. This includes: determining the weight coefficients for each of the data resource score, data quality score, computing resource score, and network resource score in a collaboration scenario based on privacy computing cooperation; and weighting the data resource score, data quality score, computing resource score, and network resource score based on the determined weight coefficients to obtain the participant's comprehensive score.
[0081] The weighting coefficient is used to characterize the degree of influence of each score in a specific cooperation scenario. The cooperation scenario refers to the combination of scenario elements such as the business type, the composition of the participants, the model training stage, and the data circulation requirements of the privacy computing cooperation.
[0082] For different collaboration scenarios, data resources, data quality, computing resources, and network resources have varying impacts on joint modeling results and collaborative inputs. Therefore, by setting their respective weight coefficients, the overall score can be matched to the actual contribution distribution. These weight coefficients can be calibrated based on historical collaboration results, model performance trends, or resource consumption characteristics, and stored in the scenario configuration table for later retrieval.
[0083] In practical implementation, the overall score of the participants can be expressed as the sum of the scores of each component weighted by a weighted coefficient. If the dimensions or numerical ranges of the scores are inconsistent, the scores can be normalized first, and then multiplied and summed according to the weighted coefficients to obtain the overall score under a unified scale. When data quality has a dominant impact on the model performance in the aforementioned collaborative scenarios, the weighted coefficient corresponding to the data quality score can be increased; when computational resource consumption or network transmission consumption is more significant in the aforementioned scenarios, the weighted coefficient corresponding to the computational resource score or network resource score can be increased accordingly.
[0084] For example, in financial risk control scenarios, the weighting coefficient of data quality score Maximum value, taking the value of 0.4, is the weighting coefficient of data resource scores in marketing scenarios. It is 0.5.
[0085] Through the above methods, the comprehensive score can incorporate data resources, data quality, computing resources, and network resources into a unified evaluation framework, and dynamically adjust the impact of each score according to different privacy computing cooperation scenarios, so that the comprehensive score of the participants is more in line with their actual input and contribution in the collaboration process.
[0086] Based on the aforementioned embodiments, further, the data implementation costs of each participant are obtained, including: for each participant, if the participant is a partner, the data implementation cost of the partner is determined based on the partner's data collection cost, data cleaning and de-identification cost, storage space rental cost, security guarantee cost, and daily operation and maintenance and disaster recovery cost.
[0087] The costs are categorized as follows: Data acquisition costs refer to the upfront investment by partners in acquiring raw business data for privacy computing collaboration, which may include API call costs and manual collection and maintenance costs; Data cleaning and de-identification costs include the expenses incurred by partners in data cleaning processes such as deduplication, removal of missing values, filtering of abnormal noise, and standardization of field formats for raw business data, as well as investments in encryption, de-identification, and manual review of sensitive fields such as mobile phone numbers, ID cards, and addresses; Storage space rental costs include expenditures on storage resources for storing raw business data, processed datasets, and intermediate files, including local server hard drive expansion purchases, cloud storage pay-as-you-go rental fees, and rental fees for dedicated distributed privacy computing storage clusters; Security costs mainly include the consumption of encryption computing power; and Daily maintenance and disaster recovery costs are the continuous operating costs to ensure the long-term stability of the dataset and its ability to participate in multi-party collaboration, including expenditures on multi-copy off-site backups and data loss emergency recovery systems.
[0088] The data implementation cost for the partner can be the sum of various costs including data acquisition cost, data cleaning and de-identification cost, storage space rental cost, security cost, and daily operation and maintenance and disaster recovery cost.
[0089] In practical implementation, partners can compile detailed cost breakdowns within a preset statistical period and categorize them according to cost item correspondence into data collection, data cleaning and de-identification, storage space leasing, security assurance, and daily operation and disaster recovery. These costs are then summed to obtain the partner's data implementation cost. When the partner's business scenario has high security requirements, higher accounting coefficients can be assigned to security assurance costs and daily operation and disaster recovery costs based on the partner's preset cost accounting rules before summarizing, forming a cost benchmark for subsequent revenue distribution. This data implementation cost reflects the actual end-to-end investment made by the partner before and after participating in privacy computing collaboration to provide data, and serves as input for subsequent comprehensive evaluation and incentive allocation.
[0090] By incorporating data acquisition, preprocessing, storage, security, and operational costs into a unified accounting system, the true input of partners can be quantified, and the distribution of benefits related to that input can be aligned with the actual contribution.
[0091] Based on the aforementioned embodiments, further, the data implementation cost of each participant is obtained, including: for each participant, if the participant is a relay party, the data implementation cost of the relay party is determined based on the relay party's security aggregation coordination and computing power cost, security assurance cost, and daily operation and maintenance and disaster recovery cost.
[0092] Among these, the relay party's core and unique costs include: security aggregation and coordination costs, and computing power costs. These encompass all the computing power required for multi-party encrypted communication, privacy-focused aggregation computing, and task scheduling. Security costs arise from the dedicated security investment required by the relay party as a trusted cross-institutional relay node to build a comprehensive isolation and protection system. Daily maintenance and disaster recovery costs cover the ongoing operational and fault recovery costs necessary to ensure the relay party stably provides networking, aggregation, and forwarding services.
[0093] In practical implementation, the following steps can be taken: First, collect data on the transit party's computing node usage time, computing power unit price, and security aggregation task volume within the predetermined statistical period to form the security aggregation coordination and computing power cost. Then, summarize the security component deployment costs, encrypted transmission costs, auditing and record-keeping costs, and security inspection costs to form the security assurance cost. Finally, summarize the operation and maintenance personnel input, system maintenance costs, storage backup costs, and disaster recovery drill and disaster recovery resource usage costs to form the daily operation and maintenance and disaster recovery cost. These costs are then summed to obtain the transit party's data implementation cost. When the transit party undertakes a large number of security aggregation tasks, a higher weight can be assigned to the security aggregation coordination and computing power costs before weighted summarization to ensure that the data implementation cost more closely reflects its actual investment.
[0094] By incorporating the costs of secure aggregation and coordination, computing power, security assurance, and daily operation and disaster recovery of the transit party into a unified accounting system, the service input of the transit party in the cooperation can be fully included in the data implementation cost. This will enable subsequent incentive evaluations to more accurately reflect their actual contributions and ensure that the cost accounting results are consistent with the resource consumption in privacy computing cooperation.
[0095] Based on the aforementioned embodiments, the method further includes: retrieving, for each participant, supporting materials that have undergone hash calculation and encryption processing from the blockchain or secret shared distributed storage; decrypting the supporting materials and verifying the decrypted supporting materials based on the hash value; if the verification passes, verifying the data implementation cost of the participant based on the decrypted supporting materials, and calculating excess revenue and contribution value based on the verified data implementation cost.
[0096] The supporting materials include project construction contracts and operating expense settlement statements related to the privacy computing cooperation provided by the participating parties, so as to provide factual evidence of the data implementation costs of the participating parties.
[0097] After retrieving the supporting materials, the encrypted text is first decrypted using the key corresponding to the participating party to obtain readable supporting material text or structured records. The decryption result is then recalculated to obtain a hash value, which is compared with the hash value recorded on the storage side. If they match, it indicates that the supporting material has not been tampered with during transmission or storage, and the verification passes, proceeding to the data implementation cost verification stage. During verification, the cost composition recorded in the supporting materials is compared with a preset cost caliber. The preset cost caliber includes a set of cost items corresponding to the participating party's role, and each cost is converted to the same cost benchmark value using a unified unit of measurement and summarized to obtain the summarized cost. If the deviation between the data implementation cost and the summarized cost is small, for example, if the absolute difference is less than the preset difference, then the data implementation cost passes the verification. Conversely, if the deviation between the data implementation cost and the aggregated cost is large, the data implementation cost fails the verification and an anomaly warning can be issued for manual verification. The data implementation cost can then be manually corrected and used for subsequent calculations. Alternatively, if the data implementation cost fails the verification, the aggregated cost can be used instead of the data implementation cost for subsequent calculations, including the calculation of excess profits, incentive values, etc.
[0098] Furthermore, when the decrypted supporting materials are compared with their corresponding hash values, if the two are found to be inconsistent, or if the supporting materials have missing pages, empty fields, abnormal timestamps, abnormal amounts, or mismatched signatures, an error message can be generated to prompt the corresponding partner to provide supporting materials again.
[0099] This mechanism, through the concatenated processing of decryption of supporting materials, hash verification, and cost verification, enables cost data to be traceable and verifiable. It also ensures that the calculation of excess revenue and contribution value is based on confirmed costs, thereby aligning the revenue distribution results with the actual input of the participants. This provides accurate cost input for subsequent calculations and improves the accuracy of incentive value assessment.
[0100] Based on the aforementioned embodiments, the method further includes: if the verification fails, the data implementation cost of the participating party is corrected based on the decrypted supporting materials, and the excess revenue and contribution value are calculated based on the corrected data implementation cost.
[0101] If the data implementation cost fails verification, the data implementation cost can be recalculated based on supporting materials. Specifically, the new data implementation cost can be formed by adding up each cost item in the supporting materials, or the corresponding item in the data implementation cost can be corrected based on the cost of the corresponding item in the supporting materials to obtain the corrected data implementation cost.
[0102] After obtaining the revised data implementation cost, it is incorporated into the total cost of cooperation. The excess revenue is calculated in conjunction with the total revenue generated by the privacy computing cooperation. Finally, the contribution value of the participating parties is calculated based on the correspondence between the revised data implementation cost and the excess revenue.
[0103] By incorporating supporting information in cases where verification fails into the correction process, the calculation of subsequent excess profits and contribution values can still be based on traceable and correctable costs, thereby ensuring that the profit distribution results are consistent with the actual input of the participants and improving the accuracy of incentive value calculation.
[0104] Based on the aforementioned embodiments, further, the model performance contribution value of each participant is obtained, including: obtaining the model performance index of each participant subset, wherein the participant subset consists of some participants in the privacy computing collaboration; if the number of participants in the privacy computing collaboration is less than or equal to a preset number, then for each participant, based on the Shapley value formula and the model performance index of each target participant subset excluding the participant, the expected value of the marginal improvement in model performance after the participant joins each target participant subset is calculated, thereby obtaining the model performance contribution value of the participant.
[0105] The preset quantity is a configurable parameter and can be 6, 8, 10, or other values.
[0106] In the specific implementation, when the number of participants does not exceed a preset number, all target participant subsets that do not contain a particular participant are traversed, and the model performance metrics corresponding to these subsets are read. The performance difference before and after the participant is added to a target participant subset is then calculated; this difference represents the marginal improvement. Subsequently, the marginal improvement of each target participant subset is weighted and averaged according to the Shapley value formula, where the weights are related to the size of the target participant subset and the probability of their combination, thus obtaining the model performance contribution value of that participant. If a participant can significantly improve model performance in multiple subsets, its corresponding expected marginal improvement is larger, and its model performance contribution value is correspondingly higher.
[0107] The above processing takes the model performance of the participating subset as input and outputs the model performance contribution value of each participating party through marginal contribution calculation, so that the model contribution evaluation can reflect the average role of the participating parties in different collaboration combinations.
[0108] Because the expected value calculation method based on Shapley value is adopted, the contribution value of each participant can correspond to its real impact on the joint modeling effect, and provide a quantitative basis for subsequent revenue distribution, so that the contribution evaluation results are more in line with the model revenue distribution in the multi-party privacy computing collaboration scenario.
[0109] In one possible implementation, the method further includes: if the number of participants in the privacy computing collaboration is greater than a preset number, then performing a random permutation operation on all participants in the privacy computing collaboration T times in a loop; after each random permutation operation, extracting a temporary subset formed by all participants before the current participant in the current permutation, obtaining the model performance index of the temporary subset and the model performance index after the current participant is added to the temporary subset; calculating the difference in the model performance index before and after the current participant is added to the temporary subset to obtain the marginal improvement in model performance of the current participant under the current permutation; and calculating the average of the marginal improvement in model performance of the current participant under each permutation obtained by the T random permutation operations to obtain the model performance contribution value of the current participant.
[0110] Where T is a positive integer, which is a configurable parameter, such as 500, 1000 or other values.
[0111] When the number of participants exceeds a preset limit, all participants are shuffled T times using a uniform random seed or an independent random source, resulting in T permutation sequences. For each permutation sequence, a temporary subset is formed by extracting the preceding participants according to the current participant's position in the sequence. The model evaluation interface in the privacy computing platform is then called to calculate the performance metrics when training or inference is performed using only the temporary subset, and the performance metrics after adding the current participant to the temporary subset. Subsequently, the marginal improvement in model performance under the current permutation is generated based on the difference between the two performance metrics. The marginal improvements obtained from the T shuffles are then arithmetically averaged to serve as the model performance contribution value for the current participant.
[0112] This approach approximates full quantum set enumeration through random permutations, achieving stable contribution evaluation results even with a large number of participants while reducing computational costs. Since the preorder set of the same participant differs in different permutations, the average marginal improvement mitigates the randomness of a single permutation, making the model's performance contribution closer to its true contribution to joint modeling. Thus, while maintaining evaluation feasibility, it reduces the computational overhead of combinatorial explosion and provides directly usable contribution parameters for subsequent calculations of revenue share and incentive values.
[0113] In one possible implementation, the revenue share of each participant is determined based on their overall score and model performance contribution value in the privacy computing collaboration. This includes: for each participant, determining their revenue share as the ratio of the product of their overall score and model performance contribution value to the cumulative value of the products of all participants' overall scores and model performance contribution values; and based on each participant's data implementation cost and the product of their revenue share and excess revenue, determining their incentive value. This includes: for each participant, determining their incentive value as the sum of their data implementation cost and the products of their revenue share and excess revenue.
[0114] In practical implementation, the overall score of each participant can be multiplied by its model performance contribution value to obtain the corresponding contribution weight. Then, this contribution weight is compared with the cumulative value of the corresponding contribution weights of all participants to obtain the revenue share of each participant. Subsequently, the revenue share of each participant is multiplied by the excess revenue to obtain the corresponding revenue allocation. This revenue allocation is then added to the data implementation cost of each participant to form the corresponding incentive value. If a participant has a high overall score and a large model performance contribution value, its revenue share will increase accordingly, and the revenue allocation in its incentive value will also increase accordingly. If a participant has a high data implementation cost, the basic compensation portion of its incentive value will be increased accordingly, thus matching the incentive result with the actual investment.
[0115] No. Incentive value for each participant The calculation relationship is as follows:
[0116]
[0117] in, The total number of participants; For excess returns; , and The first The data implementation cost, overall score, and model performance contribution of each participant. For the first The percentage of revenue for each participating party.
[0118] This approach determines the revenue share by combining a comprehensive score with the model performance contribution value, and forms an incentive value by adding the data implementation cost to the excess revenue allocation. This ensures that the excess revenue allocation reflects both the quality of participation and the cost of investment, thus making the revenue distribution result closer to the actual contribution of each participant in the privacy computing collaboration. Adopting this approach allows high-contributing participants to receive a higher share of the revenue and also provides basic compensation for high-cost investments, thereby ensuring strong consistency and executability of the incentive results in joint modeling.
[0119] In one possible implementation, information on computing resource usage, network resource usage, and data resource parameters of the participants in the privacy computing collaboration is obtained, including: monitoring the resource consumption of each participant's computing nodes, collecting computing power consumption time, memory usage capacity, and storage space capacity to obtain computing resource usage information of each participant; monitoring the transmission link and traffic operation indicators of each participant, collecting the total number of transmitted bytes, the number of routed and forwarded data packets, the security aggregation computing power consumption, and the data packet loss rate to obtain network resource usage information; and monitoring data flow resources through data resource probes, collecting the number of data entries, the number of feature dimensions, the update frequency, and data quality indicators to obtain data resource parameters.
[0120] In this embodiment, the resource usage information of computing nodes can be collected in real time by a monitoring agent deployed on the node side. The monitoring agent periodically records processor usage, memory residency, and local storage read / write status, and summarizes and outputs the computing power consumption time, memory usage capacity, and storage space capacity. Transmission link and traffic operation indicators can be obtained by the network-side acquisition module in combination with link probes, switching device logs, and transmission session statistics, thereby forming the total number of bytes transmitted, the number of routed and forwarded data packets, the security aggregation computing power consumption, and the data packet loss rate. Data resource probes can be implemented using embedded acquisition programs or bypass monitoring components to continuously track the flow status of data during the processes of acquisition, cleaning, de-identification, transmission, and storage, and output data resource parameters that can be used for subsequent resource evaluation.
[0121] By adopting the above acquisition method, resource representation information in three dimensions—computation, network, and data—can be generated simultaneously. This allows for the unified quantification of the participants' input in privacy computing collaborations, facilitating subsequent calculations of comprehensive scores, revenue share, and incentive values. Consequently, resource consumption and data value information have a consistent evaluation basis.
[0122] The incentive evaluation method provided in this application will be described below with reference to exemplary embodiments.
[0123] Figure 2 A framework diagram of a privacy computing collaboration system provided for this application, such as... Figure 2 As shown, the participants in the privacy computing collaboration include partner A, partner B, partner C, and intermediary D. For example, it could also include partner N, where partners A, B, and C are three partners whose networks are interconnected, while partner N is only interconnected with intermediary D.
[0124] Figure 2 The scenarios shown are merely examples and should not be considered as limiting the application scenarios of this application. This application also supports collaborative scenarios involving more partners.
[0125] See also Figure 2 Each participating party is equipped with a data circulation component. The data circulation component mainly includes a contribution value evaluation and measurement module, an incentive quantification calculation module, and a transmission networking module. If the relay party D also undertakes the role of data routing and forwarding, its deployed data circulation component also includes a routing and forwarding module, as shown in the schematic diagram of the internal functional modules of the data circulation component of the central node, relay party D.
[0126] Before incentive evaluation, resource networking is required. During platform operation, participating organizations will collaborate using the data circulation components within their respective privacy computing platforms. The transmission networking module within the data circulation components is crucial for this collaborative networking. This module assigns NodeIDs to registered participants. During registration, participants must submit their organizational information, platform information, and information on platform resources (such as computing power and memory) for subsequent multi-party data collaboration and incentive value calculation. Depending on the collaboration needs, the transmission networking module can be configured with strategies such as resource probes to statistically verify the resources registered by each party during collaboration based on the transmission networking module.
[0127] In the field of financial risk control, assuming that by Figure 2Partners A, B, C, and relay party D jointly conduct privacy-preserving computation. Partner A can be a bank, partner B can be a telecom operator, and partner C can be an e-commerce platform. Multiple institutions collaborate on modeling using privacy-preserving computation technology to improve the accuracy of credit risk prediction models. The data from each institution is distributed across heterogeneous privacy-preserving computation platforms, with differences in data quality, computing power consumption, and network resources. Partners A, B, C, and D jointly conduct credit risk prediction modeling. Partner A provides 10 million credit data points, partner B provides 5 million user behavior data points, partner C provides 8 million consumption data points, and relay party D is responsible for cross-network routing and secure aggregation. The goal of the joint modeling is to increase the AUC from 0.75 to 0.92, and to distribute rewards based on contribution, i.e., incentive value.
[0128] In one possible implementation, partners A, B, C, and relay party D register and join the network via the transmission networking module, submitting their respective resource information, including computing power, storage, and network bandwidth. Resource probes are deployed to collect resource consumption data in real time. For each participant in the privacy computing collaboration, resource input and data resource status are collected during the participation process. This includes monitoring the resource usage of each participant's computing nodes, collecting computing power consumption time, memory usage, and storage space capacity to obtain computing resource usage information. For example, partner A's probe detected 100 hours of GPU computing power consumption, combined with 50GB of memory, to obtain a computing resource score. =100 hours × 50 GB = 5000 (hours·GB). Simultaneously, the transmission links and traffic operation indicators of each participant are monitored, and the total number of bytes transmitted, the number of data packets forwarded, the security aggregation computing power consumption, and the data packet loss rate are collected to obtain network resource usage information; the probe of partner B recorded its network bandwidth usage as 30 GB / s, and the probe of relay party D collected its number of data packets forwarded as 2 million packets.
[0129] Data resource parameters are obtained by monitoring data flow resources through data resource probes, collecting data count, feature dimension count, update frequency, and data quality indicators. In Partner A's data resource parameters, there are 10 million data records and 50 feature dimensions. Based on these parameters, Partner A's data resource score is determined. =10 million records × 50 fields = 50,000 (10,000 records·fields); Partner B provides 5 million user behavior data records, and Partner C provides 8 million consumption data records. Each partner has its own data record count, feature dimension count, update frequency, and data quality metrics. A data resource score is determined based on the data record count, feature dimension count, and update frequency. A data quality score is determined based on the data quality metrics. Partner A's data quality score is... =90.5 points, with 95% completeness, 90% accuracy, and 85% consistency. The computing resource score is determined based on computing resource usage information, the network resource score is determined based on network resource usage information, and the weighting coefficients for the data resource score, data quality score, computing resource score, and network resource score are determined in conjunction with the cooperation scenario of privacy computing collaboration; in the financial risk control scenario, the weight of the data quality score... =0.4, the final resource assessment value corresponding to partner A is the comprehensive score. Calculate using the following formula: 0.2×50000+0.2×5000+0.2×2500+0.4×90.5, to obtain... =11536.2.
[0130] When acquiring cost data, each participant's role is considered and processed accordingly. When a participant is a partner, the data implementation cost for that participant is determined based on data acquisition costs, data cleaning and de-identification costs, storage space rental costs, security costs, and daily operation and disaster recovery costs. For partner A, as a bank, its data implementation cost is... =2.1 million yuan, which includes costs for data collection, cleaning, and storage. When a participant is a relay party, the data implementation cost for that participant is determined based on the costs of secure aggregation and coordination, computing power, security assurance, and daily operation and disaster recovery. For relay party D, as a relay party, its data implementation cost is... The amount is 300,000 yuan. Subsequently, supporting materials from each participant, processed by hashing and encryption, are retrieved from the blockchain or a secret shared distributed storage. These supporting materials are then decrypted, and the hash values are used to verify them. If the verification passes, the data implementation cost is verified based on the decrypted supporting materials, and this verified cost is used in subsequent calculations. If the verification fails, the data implementation cost is corrected based on the decrypted supporting materials, and the corrected cost is used in subsequent calculations.
[0131] The model performance contribution value is calculated using the model performance metrics of the participating subsets. First, the model performance metrics corresponding to each participating subset, consisting of a subset of participating parties, are obtained. When the number of participating parties in the privacy computing collaboration is less than or equal to a preset number, for each participating party, based on the Shapley value formula and combined with the model performance metrics of each target participating subset excluding that participating party, the expected marginal improvement in model performance after that participating party joins each target participating subset is calculated, yielding the model performance contribution value for that participating party. When the number of participating parties in the privacy computing collaboration is greater than the preset number, a random permutation operation is performed T times for all participating parties. Each time a random permutation operation is performed, all participating parties preceding the current participating party in the current permutation are extracted to form a temporary subset. The model performance metrics of the temporary subset and the model performance metrics after the current participating party joins the temporary subset are obtained respectively. The difference in model performance metrics before and after the current participating party joins the temporary subset is calculated, yielding the marginal improvement in model performance for the current participating party under that permutation. The average of the marginal improvement values obtained from the T random permutations is then used to obtain the model performance contribution value for the current participating party.
[0132] In this financial risk control example, the Shapley value is calculated using the Monte Carlo approximation, and the model performance contribution value of partner A is... =0.1667, the marginal contribution of the intermediate party D to the model convergence speed is calculated using the Shapley value. =0.0567, the Monte Carlo approximation method reduces the computational complexity of the Shapley value from O(2^3) to O(2^3). n (It has been reduced to an acceptable level.)
[0133] After the overall scores, data implementation costs, and model performance contributions of each participant have been determined, excess revenue is calculated based on each participant's model performance contribution and data implementation cost. Then, the revenue share for each participant is calculated as the ratio of the product of that participant's overall score and model performance contribution to the cumulative product of the overall scores and model performance contribution values of all participants. In this example, the total revenue is 6 million yuan, the total cost is 5.31 million yuan, and the technology premium pool... =690,000 yuan, The excess profit is then determined based on the data implementation costs of each participant, and the sum of the products of each participant's profit share and the excess profit. The incentive value for each participant is calculated by adding the participant's data implementation cost to the product of their profit share and the excess profit, using the following formula:
[0134]
[0135] The incentive value for partner A is calculated using the aforementioned formula. for: Ten thousand yuan.
[0136] The routing, forwarding, and security aggregation resource consumption of relay party D are included in the evaluation, and its network resource score is calculated accordingly. =2 million packets × 0.1 GB × 5 ms = 1000 (million packets·GB·ms), corresponding to the overall score The value is 272, and the excitation value of the transferor D is calculated according to the aforementioned formula. for: Ten thousand yuan.
[0137] In accordance with the methods provided in the foregoing embodiments, this application also provides a corresponding apparatus. Figure 3 A schematic diagram of the privacy computing participant incentive evaluation device provided in this application is shown below. Figure 3 As shown, the privacy computing participant incentive evaluation device 30 provided in this embodiment includes: a data acquisition module 310, used to acquire computing resource usage information, network resource usage information, and data resource parameters of each participant in the privacy computing collaboration; the data resource parameters include the number of data entries, the number of feature dimensions, the update frequency, and data quality indicators; a comprehensive score determination module 320, used to determine the comprehensive score of each participant based on the computing resource usage information, network resource usage information, and data resource parameters; and a cost acquisition module 330, used to acquire the data implementation cost and model performance contribution value of each participant; wherein, the data implementation cost... The implementation cost is used to characterize the total implementation cost invested by the participants in participating in privacy computing; the model performance contribution value is used to characterize the marginal improvement in model performance brought about by the participation of the participants; the excess revenue calculation module 340 is used to calculate the excess revenue based on the model performance contribution value and data implementation cost of each participant in the privacy computing collaboration; the revenue share determination module 350 is used to determine the revenue share of each participant based on the comprehensive score and model performance contribution value of each participant in the privacy computing collaboration; the incentive value determination module 360 is used to determine the incentive value of each participant based on the data implementation cost of each participant and the product of the revenue share and excess revenue of each participant.
[0138] In one possible embodiment, the comprehensive score determination module 320 includes: a first score determination unit, used to determine a data resource score based on the number of data entries, the number of feature dimensions, and the update frequency; a second score determination unit, used to determine a data quality score based on data quality indicators; a third score determination unit, used to determine a participant's computing resource score based on the participant's computing resource usage information; a fourth score determination unit, used to determine a participant's network resource score based on the participant's network resource usage information; and a comprehensive score determination unit, used to determine a participant's comprehensive score based on the participant's data resource score, data quality score, computing resource score, and network resource score.
[0139] In one possible embodiment, the comprehensive score determination unit is specifically used for: determining the weight coefficients of data resource score, data quality score, computing resource score, and network resource score in a cooperation scenario based on privacy computing cooperation; and weighting the data resource score, data quality score, computing resource score, and network resource score based on the determined weight coefficients to obtain the comprehensive score of the participants.
[0140] In one possible embodiment, when acquiring the data implementation costs of each participant, the cost acquisition module 330 is specifically used to: for each participant, if the participant is a partner, determine the partner's data implementation costs based on the partner's data acquisition costs, data cleaning and de-identification costs, storage space rental costs, security assurance costs, and daily operation and maintenance and disaster recovery costs.
[0141] In one possible embodiment, when acquiring the data implementation cost of each participant, the cost acquisition module 330 is further configured to: for each participant, if the participant is a relay party, determine the data implementation cost of the relay party based on the relay party's security aggregation coordination and computing power cost, security assurance cost, and daily operation and maintenance and disaster recovery cost.
[0142] In one possible embodiment, the cost acquisition module 330 is further configured to: retrieve, for each participant, supporting materials that have undergone hashing and encryption processing from the blockchain or a secret shared distributed storage; decrypt the supporting materials and verify the decrypted supporting materials based on the hash value; if the verification passes, verify the data implementation cost of the participant based on the decrypted supporting materials, and calculate the excess revenue and contribution value based on the verified data implementation cost.
[0143] In one possible embodiment, the device further includes a cost correction module for: correcting the data implementation cost of the participating party based on the decrypted supporting materials if the verification fails, so as to calculate excess revenue and contribution value based on the corrected data implementation cost.
[0144] In one possible embodiment, when the cost acquisition module 330 acquires the model performance contribution value of each participant, it is specifically used to: acquire the model performance index of each participant subset, wherein the participant subset consists of some participants in the privacy computing collaboration; if the number of participants in the privacy computing collaboration is less than or equal to a preset number, then for each participant, based on the Shapley value formula and the model performance index of each target participant subset excluding the participant, calculate the expected value of the marginal improvement in model performance after the participant joins each target participant subset, and obtain the model performance contribution value of the participant.
[0145] In one possible embodiment, the cost acquisition module 330, when acquiring the model performance contribution value of each participant, is further configured to: if the number of participants in the privacy computing collaboration is greater than a preset number, perform a random permutation operation on all participants in the privacy computing collaboration T times in a loop; each time a random permutation operation is performed, extract a temporary subset formed by all participants before the current participant in the current permutation, acquire the model performance index of the temporary subset, and the model performance index after the current participant is added to the temporary subset; calculate the difference in the model performance index before and after the current participant is added to the temporary subset to obtain the marginal improvement in model performance of the current participant under the current permutation; calculate the average value of the marginal improvement in model performance of the current participant under each permutation obtained by the T random permutation operations to obtain the model performance contribution value of the current participant.
[0146] In one possible embodiment, the revenue share determination module 350 is specifically used to: determine the revenue share of each participant as the ratio of the product of the participant's overall score and the model performance contribution value to the cumulative value of the product of the overall score and the model performance contribution value of each participant.
[0147] In one possible embodiment, the incentive value determination module 360 is specifically used to: determine the incentive value for each participant as the sum of the product of the participant's data implementation cost, the participant's revenue share, and the excess revenue.
[0148] In one possible embodiment, the data acquisition module 310 is specifically used to: monitor the resource consumption of each participant's computing nodes, collect computing power consumption time, memory usage capacity and storage space capacity to obtain computing resource usage information of each participant; monitor the transmission link and traffic operation indicators of each participant, collect the total number of transmitted bytes, the number of routed and forwarded data packets, the computing power consumption of security aggregation and the data packet loss rate to obtain network resource usage information; and monitor data flow resources through data resource probes, collect the number of data entries, feature dimensions, update frequency and data quality indicators to obtain data resource parameters.
[0149] The privacy computing participant incentive evaluation device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.
[0150] Figure 4 A schematic diagram of the structure of the electronic device provided in this application. Figure 4 As shown, the electronic device 40 provided in this embodiment includes at least one processor 401 and a memory 402. Optionally, the electronic device 40 further includes a communication component 403. The processor 401, memory 402, and communication component 403 are connected via a bus 404.
[0151] In a specific implementation, at least one processor 401 executes computer execution instructions stored in memory 402, causing at least one processor 401 to perform the above-described method.
[0152] The specific implementation process of processor 401 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.
[0153] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.
[0154] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.
[0155] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.
[0156] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.
[0157] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.
[0158] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static 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), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.
[0159] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.
[0160] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.
[0161] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0162] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0163] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0164] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.
[0165] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.
Claims
1. A method for evaluating incentives for participants in privacy-preserving computation, characterized in that, include: For each participant in the privacy computing collaboration, obtain the computing resource usage information, network resource usage information, and data resource parameters of the participants in the privacy computing collaboration; the data resource parameters include the number of data entries, the number of feature dimensions, the update frequency, and data quality indicators. Based on the participants' computing resource usage information, network resource usage information, and data resource parameters, the participants' comprehensive score is determined. Obtain the data implementation cost and model performance contribution value of each participating party; wherein, the data implementation cost is used to characterize the total implementation cost invested by the participating party in participating in privacy computation, and the model performance contribution value is used to characterize the marginal improvement in model performance brought about by the participation of the participating party; The excess revenue is calculated based on the model performance contribution and data implementation cost of each participant in the privacy computing collaboration. Based on the comprehensive scores and model performance contributions of each participant in the privacy computing collaboration, the revenue share of each participant is determined. The incentive value for each participant is determined based on the data implementation cost of each participant and the product of the participant's share of the revenue and the excess revenue.
2. The method according to claim 1, characterized in that, The determination of a participant's comprehensive score based on the participant's computing resource usage information, network resource usage information, and data resource parameters includes: The data resource score is determined based on the number of data entries, the number of feature dimensions, and the update frequency. Based on the aforementioned data quality indicators, a data quality score is determined; Based on the computing resource usage information of the participants, the computing resource score of the participants is determined; Based on the network resource usage information of the participants, the network resource score of the participants is determined; The comprehensive score of each participant is determined based on their data resource score, data quality score, computing resource score, and network resource score.
3. The method according to claim 2, characterized in that, The determination of a participant's overall score based on the participant's data resource score, data quality score, computing resource score, and network resource score includes: Based on the cooperation scenario of the privacy computing cooperation, the weight coefficients of the data resource score, data quality score, computing resource score and network resource score are determined respectively. Based on the determined weighting coefficients, the data resource score, data quality score, computing resource score, and network resource score are weighted to obtain the comprehensive score of the participating party.
4. The method according to claim 1, characterized in that, The implementation costs of acquiring data from each of the aforementioned participants include: For each of the aforementioned participants, if the participant is a partner, the data implementation cost of the partner is determined based on the partner's data acquisition cost, data cleaning and de-identification cost, storage space rental cost, security guarantee cost, and daily operation and maintenance and disaster recovery cost.
5. The method according to claim 1, characterized in that, The implementation costs of acquiring data from each of the aforementioned participants include: For each of the aforementioned participants, if the participant is a relay party, the data implementation cost of the relay party is determined based on the relay party's security aggregation and coordination costs, computing power costs, security assurance costs, and daily operation and maintenance and disaster recovery costs.
6. The method according to claim 1, characterized in that, The method further includes: For each of the aforementioned participants, retrieve the supporting materials of the participants that have undergone hash calculation and encryption processing from the blockchain or secret shared distributed storage; The supporting evidence is decrypted, and the decrypted supporting evidence is verified based on the hash value. If the verification passes, the data implementation cost of the participating party is verified based on the decrypted supporting materials, and the excess revenue and contribution value are calculated based on the verified data implementation cost.
7. The method according to claim 6, characterized in that, The method further includes: If the verification fails, the data implementation cost of the participating party will be adjusted based on the decrypted supporting materials, and the excess revenue and contribution value will be calculated based on the adjusted data implementation cost.
8. The method according to claim 1, characterized in that, Obtain the model performance contribution value of each of the aforementioned participants, including: Obtain model performance metrics for each subset of participating parties, wherein the subset of participating parties consists of some of the participants in the privacy computing collaboration; If the number of participants in the privacy computing collaboration is less than or equal to a preset number, then for each participant, based on the Shapley value formula and the model performance index of each target participant subset excluding the participant, the expected value of the marginal improvement in model performance after the participant is added to each target participant subset is calculated, and the model performance contribution value of the participant is obtained.
9. The method according to claim 8, characterized in that, The method further includes: If the number of participants in the privacy computing collaboration is greater than the preset number, then the random permutation operation of all participants in the privacy computing collaboration is performed T times in a loop. Each time a random permutation operation is performed, a temporary subset is extracted from all participants before the current participant in the current permutation. The model performance metrics of the temporary subset and the model performance metrics after the current participant is added to the temporary subset are obtained respectively. Calculate the difference in model performance metrics before and after the current participant is added to the temporary subset, and obtain the marginal improvement in model performance of the current participant under the current arrangement; Calculate the average marginal improvement in model performance of the current participant under each of the T random permutations to obtain the model performance contribution value of the current participant.
10. The method according to any one of claims 1-9, characterized in that, The revenue share of each participant in the privacy-preserving computing collaboration is determined by combining their comprehensive scores and model performance contributions, including: For each of the aforementioned participants, the revenue allocation for each participant is determined as the ratio of the product of the participant's overall score and the model performance contribution value to the cumulative value of the product of the overall score and the model performance contribution value for each participant. The determination of the incentive value for each participant, based on the data implementation cost of each participant and the product of the participant's share of revenue and the excess revenue, includes: For each of the aforementioned participants, the incentive value for that participant is determined to be the sum of the products of the participant's data implementation cost, the participant's share of the revenue, and the excess revenue.
11. The method according to any one of claims 1-9, characterized in that, The acquisition of the participants' computing resource usage information, network resource usage information, and data resource parameters in the privacy computing collaboration includes: Monitor the resource consumption of each participant's computing node, collect computing power consumption time, memory usage capacity and storage space capacity, and obtain computing resource usage information of each participant; Monitor the transmission links and traffic operation indicators of each participant, collect the total number of transmitted bytes, the number of routing and forwarding data packets, the consumption of security aggregation computing power, and the data packet loss rate to obtain network resource usage information; By monitoring data flow resources through data resource probes, data resource parameters are obtained by collecting the number of data entries, feature dimensions, update frequency, and data quality indicators.
12. A privacy-preserving computation participant incentive evaluation device, characterized in that, include: The data acquisition module is used to acquire information on the computing resource usage, network resource usage, and data resource parameters of each participant in the privacy computing collaboration. The data resource parameters include the number of data entries, the number of feature dimensions, the update frequency, and data quality indicators. The comprehensive score determination module is used to determine the comprehensive score of the participants based on their computing resource usage information, network resource usage information, and data resource parameters. The cost acquisition module is used to acquire the data implementation cost and model performance contribution value of each participant; wherein, the data implementation cost is used to characterize the total implementation cost invested by the participant in participating in privacy computing, and the model performance contribution value is used to characterize the marginal improvement in model performance brought about by the participation of the participant; The excess profit calculation module is used to calculate excess profits based on the model performance contribution value and data implementation cost of each participant in the privacy computing collaboration. The revenue share determination module is used to determine the revenue share of each participant based on the comprehensive score and model performance contribution value of each participant in the privacy computing collaboration. The incentive value determination module is used to determine the incentive value of each participant based on the data implementation cost of each participant and the product of the revenue share of each participant and the excess revenue.
13. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-11.
14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-11.
15. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-11.
